Changelog

What's new in each published TRW release. Every entry lists the package versions it shipped and leads with what you can now do. Steps you need to take start with “Upgrade”, and routine fixes get one line at the end.

Latest versions

TRW does not have one product-wide version number. The framework protocol (the phases, gates and evidence rules agents follow) is versioned on its own, and each package carries its own semantic version. The two move independently: a new trw-mcp major is a new major version of the MCP server package, not a new version of TRW as a whole, and the protocol can change without any package changing its major version. Each release below is labelled with the exact package versions it shipped.

v27.6

Framework protocol

9.5.1

trw-mcp

5.4.0

trw-memory

0.34.0

platform

These are the in-repo development versions of each package, and they can sit several releases ahead of what is published. The installer (install.sh) and pip resolve the most recent release published to PyPI, which is what you actually get. The release timeline below covers published releases only. Run pip index versions trw-mcp to see what is installable right now.

Releases

2026-10-11trw-mcp 9.5.1trw-memory 5.4.0

Recovery after compaction keeps what you recorded, and the installer refuses before it fetches

Upgrade: rerun the installer, or pip install -U trw-mcp, then reconnect your MCP clients and run trw-mcp update-project for the new PreCompact hook. trw-memory stays at 5.4.0.
A pre-compaction checkpoint keeps the run and the message you give it, and the PreCompact hook keeps the directive and context anchor that checkpoint recorded.
The installer has no canary channel. The served installer refuses an inherited TRW_CHANNEL other than stable before it downloads anything, and a failed health check names its own fix.
A first install adds no status line unless you ask for one, and a trw-mcp dispatch fan-out waits for a free slot as long as a single dispatch does.
Session start says when trw_learn cannot check new learnings for duplicates because the memory store needs a re-embed.
Update, doctor, uninstall and a kept REVIEW.md say what happened and name a fix that exists.

2026-10-10trw-mcp 9.5.0trw-memory 5.4.0

Claims that respect your checkout, and installs that finish outside git

Upgrade: rerun the installer, or pip install -U trw-mcp, then reconnect your MCP clients. trw-memory 5.4.0 comes with it. An installation left on 9.4.0 keeps recalling and writing; its maintain-verify needs the upgrade to read a new count.
A learning now records the commit its assertions were written at. A checkout that git shows to be behind that commit holds the claim as unknown instead of recording a failure.
Installing for two or more clients can complete outside a git repository, and the install log keeps the last frame of a menu that redraws in place.
trw-mcp telemetry channel-stats and trw-mcp handoff new answer about the project TRW is bound to, not the directory they were started in.
When trw_build_check cannot bind a record to your working tree, a nested repository with no commit is named, and any other git failure logs the step that failed.
An installation that meets a memory daemon of a later major version is told to upgrade, where it used to be told to stop the daemon.

2026-10-09trw-mcp 9.4.0trw-memory 5.3.0

Tools that read your project wherever the server runs, a delivery gate that follows your own checkout, and safer updates

Upgrade: rerun the installer, or pip install -U trw-mcp, then reconnect your MCP clients. Codex users: run trw-mcp update-project so trw_code works in sessions that never ask for approval.
trw_review, trw_code hints, trw_build_check, trw_dispatch and trw_deliver now read your project's repository even when the server was started in another directory.
The delivery gate counts the requirement status changes your checkout made, when it can tell which those are, and trw_review no longer records a pass over findings it ignored.
update-project keeps the registration of a hook you edited, its dry run matches the real run for a settings file with uncommitted changes, and it tells you when an update leaves an enrolled project blocked.
trw-memory: recall keeps working after you upgrade trw-memory in the Python that serves your store, and MemoryClient.store() accepts up to three code anchors per memory.
The code risk report reads what the current optional trw-distill writes. Settings that never had an effect (nudge_budget_chars, density_weight_*, reversion_rate_*) are retired; a leftover value gets a warning.

2026-10-08trw-mcp 9.3.0trw-memory 5.2.4

Handoff constraints you can verify, pre-edit hints in worktrees, and patched dependencies

Upgrade: rerun the installer, or pip install -U trw-mcp, then reconnect your MCP clients. Installs now pull PyJWT 2.15.0 or later.
trw-mcp handoff new --constraint-from PATH#L<first>-L<last> copies a constraint from its source file, and handoff check --strict accepts only positive evidence.
trw-mcp handoff brief turns a handoff record into read-only briefs for helper agents, and brief-check verifies every line they quote.
With the optional trw-distill package, the pre-edit hint now works from linked git worktrees and says once per session when it has nothing to read.
The pre-edit hook confirms which repository a file belongs to before reading anything, and every step it takes has a time limit.
trw-memory: the embeddings extra requires sentence-transformers 5.6.0 or later. hint_sidecar_auto_refresh_enabled is retired; use hint_sidecar_refresh_enabled.

2026-10-08trw-mcp 9.2.2

Answers that stick, update-project --rerender, a quieter status line, and sealed handoffs

Upgrade: rerun the installer, or pip install -U trw-mcp, then reconnect your MCP clients.
A remembered “no” now holds on unattended installs, and a failed prompt is never saved as an answer.
update-project --rerender PATH refreshes a TRW file you changed, after backing up your copy; a file you deleted comes back on the next update.
--ide all updates the clients your project already uses instead of adding every client.
The status line shows only what needs attention, and local commands find the right project from any subfolder.
Handoff records follow AHR 1.0-rc.2, and session start lists the ones waiting for you.

2026-10-07trw-mcp 9.2.1

An installer that remembers your answers, config unset, and Claude Haiku 5.5

Upgrade: rerun the installer, or pip install -U trw-mcp, then reconnect your MCP clients.
The installer asks each question once and keeps your answers, even through an aborted run; --reconfigure asks again.
An update refreshes every client in one pass, names each kept file once and ends with one status line.
trw-mcp config unset removes a key, retired ones included, and keeps every other line of your file as it was.
Claude Haiku 5.5 is priced, keeps its effort level in dispatch, and is the fast LLM route.
Your own skill symlinks no longer block an update, and a retired skill is removed whole or kept whole.

2026-10-07trw-mcp 9.2.0trw-memory 5.2.3

Dispatch setup in the installer, safe config edits, prompts that never wait on recall

Upgrade: rerun the installer, or pip install -U trw-mcp trw-memory, then reconnect your MCP clients.
The installer can turn on trw_dispatch and pin a model and effort per client; trw-mcp config dispatch does the same later, and headless installs take flags.
trw-mcp config set changes one setting safely: it validates the value exactly as TRW loads it, refuses secrets, and keeps your file’s quotes and layout.
The prompt hook runs inside a time budget and skips recall instead of stalling your client, and a recalled learning counts as shown only once it reaches you.
Installs and updates name each changed file once with its reason, end with one status line, and report a doctor timeout as incomplete rather than FAILED.
backup restore never touches another project’s data, and doctor suggests deleting only files TRW provably shipped and retired.
A deduplicated memory keeps its evidence level, so a merge can no longer raise a learning’s confidence beyond what supports it.

2026-10-06trw-mcp 9.1.1trw-memory 5.2.2

Handoffs another session can trust, one setup per computer, safer installs

Upgrade: rerun the installer, or pip install -U trw-mcp trw-memory, then reconnect your MCP clients.
New trw-handoff skill for Claude Code and Codex: a sealed handoff record with labelled claims and verbatim constraints, and a receive step that checks the record before acting.
Sign in once per computer, and configure trw_assess once: new project installs reuse ~/.trw instead of asking again.
An install for Claude Code plus another client keeps its Claude Code hooks, and doctor fails loudly if a registered TRW hook is missing.
Updating a project with an older channel manifest migrates it instead of rolling the whole update back.
New installs default to Claude Code, Codex and Antigravity CLI; opt-in pre-approval of TRW's own Codex hooks.
The installer keeps your API key out of process listings and writes credentials owner-only from the first byte.

2026-10-05trw-mcp 9.1.0

TRW progress in your status line, a fan-out cap, smoother upgrades

Upgrade: rerun the installer, or pip install -U trw-mcp, then reconnect your MCP clients.
Claude Code shows TRW progress in its status line, and trw-mcp local status --json gives any client the same read-only snapshot.
An opt-in cap on how many dispatch children run at once across your sessions, with an effort requirement you can turn on.
Learning and recall keep working after an upgrade when the memory daemon is stricter than the server, and a weaker daemon is still refused.
Upgrades name every file they remove, keep your telemetry answer, and leave a CLAUDE.md that already imports AGENTS.md alone.
The install run log holds the whole install, and doctor names MCP servers still running code from before the upgrade.

2026-10-03trw-mcp 9.0.1trw-memory 5.2.1

Safer updates and uninstalls, sync that keeps moving, tighter label checks

Upgrade: pip install -U trw-mcp trw-memory, then restart the MCP server and the memory daemon.
update-project keeps a local edit to FRAMEWORK.md or AARE-F-FRAMEWORK.md that git is told to ignore, and a TRW marker shown inside a code fence no longer rolls the whole update back.
uninstall no longer fails when a background codebase-map build is still writing; it waits for the writer to finish.
A backlog of held learnings no longer stops newer learnings from syncing: the push pages past held rows instead of re-reading them.
A long-running MCP server keeps reading run journals however many projects it serves, instead of failing once its open-file cache fills.
Before you edit a module that many files import, the Claude Code pre-edit hint names how many depend on it and how to run their tests.
Confidentiality labels match a tag however it is spaced, cased or encoded, and a queued publish retry checks the label again before sending.

2026-10-02trw-mcp 9.0.0trw-memory 5.2.0

Stable 9.0.0: recall queries stay local unless team sync is on

A recall query leaves your machine only when team sync is turned on, and it is redacted first. Pattern redaction cannot catch internal names typed into a query.
/trw-prd-new is removed (use /trw-prd-ready), the self_review_blocking setting is removed, and the deprecated intent-hook settings are now ignored.

2026-10-01trw-mcp 8.1.8trw-memory 5.1.8

Leaner deliver step and clearer skill cleanup reports

trw_deliver no longer runs a tier sweep over old learning files that nothing read, so delivery has one less step to wait on.
trw-mcp doctor now judges each client skill folder the same way an update does, so a copy the update kept is always named.
Settings retired in this release log one warning if your config still sets them.

2026-10-01trw-mcp 8.1.7trw-memory 5.1.7

Updates retire old skills safely and say where files went

An update retires a TRW skill or agent only when TRW no longer ships it, and only after proving TRW wrote those exact bytes. Files you edited are kept and reported.
update-project now names the folder in .trw/trash that holds each file it moved, so you can get it back.
A stale deletion record can no longer remove a file of yours that TRW never writes.
Plus fixes for the interactive installer looking hung after a failed step.

2026-10-01trw-mcp 8.1.6trw-memory 5.1.6

Recall flags stale learnings, and edited hooks survive updates

trw_recall now marks a learning whose last stored check failed as stale, so an agent knows to re-verify it before acting on it.
Updates keep Copilot hook scripts you edited, and a hook no longer runs with one old and one new library.
Plus installer fixes. Settings retired in this release log one warning if your config still sets them.

2026-10-01trw-mcp 8.1.5trw-memory 5.1.5

Read-back handoffs and automatic memory snapshots

In a formation, a handoff now moves responsibility only after the receiving agent reads it back. Each step goes into one append-only, hash-chained log, and trw_inbox shows who owns each handoff and where it stands.
memory_maintain now takes rolling daily and weekly restore points of your memory store automatically.
Upgrade: the formation mailbox has a new format. If you use formations, stop every member and run trw-mcp formation comms-upgrade.
Plus fixes for the installer's health check, memory daemon upgrades and hook updates.

2026-10-01trw-mcp 8.1.4trw-memory 5.1.4

Safer backups and private learnings

backup create no longer uploads a store that holds other projects' data. If the store contains any namespace besides the current project's, the remote upload is refused and nothing is sent.
New personal (user-tier) learnings live only in the memory store and no longer get a copy in the project's tracked .trw/learnings/ folder. trw-mcp doctor flags older copies so you can delete them, and the stdio memory server refuses to serve the machine-wide user store.
Plus clearer installer and upgrade messages.

2026-09-30trw-mcp 8.1.2–8.1.3trw-memory 5.1.2–5.1.3

Shared servers that replace themselves, and safer restores

When you upgrade trw-mcp or trw-memory, a shared server now replaces itself, with no client restart and no /mcp reconnect. If the new server fails to boot, the old one keeps serving. Set shared_mcp: {auto_swap: false} to turn it off.
backup restore now checks the archive (checksum, SQLite header and integrity) before it touches your live store, so a corrupt or truncated file changes nothing.
Servers start quietly, with no banner and no network call.
Plus clearer tool errors, and upgrades now name every file they kept or moved.

2026-09-30trw-mcp 8.1.1trw-memory 5.1.1

Instruction drift checks in CI, and opt-in tracing

trw-mcp instructions check exits 1 when an instruction file would change, so a CI step can catch instruction drift.
OpenTelemetry tracing is available and off by default: set TRW_OTEL_ENABLED and install trw-mcp[otel]. A scrubbing exporter removes content before anything leaves the process.
Feedback you submit is saved to a local outbox first. A failed send is kept for a retry, and the same report is not sent twice.
Cursor and Codex sessions now record their file edits the way Claude Code does, so trw_deliver asks for a build after code changes. Codex edits made through the shell are not seen yet. The memory write gate also refuses more secret formats, including JWTs and PEM private keys.
Upgrade: trw-mcp 8.1.1 requires trw-memory 5.1.0 or later. Run trw-mcp update-project to refresh the Cursor and Codex hooks. 8.1.0 was not published separately; its changes ship here.

2026-09-29trw-mcp 8.0.1trw-memory 5.0.1

A hook security fix, and instruction files that keep your lines

Hooks no longer source a hook-env file that is symlinked, foreign-owned, writable by others or tracked in git, so a checked-out branch can't use one to run commands. Run trw-mcp update-project to regenerate a legitimate one.
Removing an old TRW block no longer deletes your own lines, and an instruction file that only points elsewhere (such as an AGENTS.md holding just @CLAUDE.md) is left exactly as it is.
Plus fixes for uninstall on Linux and WSL2, doctor on a fresh project, and memory consolidation.

2026-09-29trw-mcp 8.0.0trw-memory 5.0.0

Half the instruction context, and evidence you can hand to a reviewer

Instruction context is about half the size on every client, with every per-turn rule kept. Claude Code goes from 3,421 to 1,713 tokens, Codex from 3,269 to 1,527, OpenCode from 2,937 to 1,452 and Cursor from 4,018 to 2,309.
trw-mcp run evidence-pack exports what a run recorded (requirements, evidence receipts, decisions and the delivery verdict) as one redacted, tamper-evident file. Review a pack before sharing it outside your team.
trw_deliver reads each PRD's approved acceptance criteria back from git history, so a weakened criterion or lost evidence shows up at delivery. Safety-critical PRDs block until the change is recorded as an amendment.
Handoffs now close: a peer accepts a request and reports completion with a pointer to what to read next, and only the requester records it complete. Updates also leave your hand-edited files in place.
Upgrade: Python 3.11 or newer is required. Install trw-memory 5.0.0 first, then trw-mcp 8.0.0, and run trw-mcp update-project. If you use formations, run trw-mcp formation comms-upgrade once per mailbox. trw-memory settings now need the prefix (MEMORY_<SETTING>), and AGENTS.md links .trw/INSTRUCTIONS.md rather than embedding it.

2026-09-26trw-mcp 7.0.1trw-memory 4.0.1

Dry runs that write nothing, and honest memory checks

Upgrade: run pip install -U trw-mcp trw-memory, then restart the MCP server and the memory daemon.
trw-mcp instructions sync --dry-run now writes nothing (in 7.0.0 a dry run still rewrote several instruction files), and memory_verify reports error or skipped when a sweep could not finish cleanly.
Plus fixes for scoped code indexing and memory imports.

2026-09-25trw-mcp 7.0.0trw-memory 4.0.0

Fewer tools, better tool picks and a safer memory store

51 MCP tools became 15, with 13 on by default. In our tool-selection eval, the first TRW call was the right one 73% of the time on Claude Code, up from 30%, and 80% on Codex, up from 0%.
Two data-integrity bugs from trw-memory 2.x and 3.x are fixed: one could corrupt a store shared by two processes, the other could lose the first writes to a new store. If you ran an earlier release, run pragma integrity_check on your stores and look for memory.db.corrupt.*.bak files.
trw_session_start p95 fell from 718–856 ms to 390–462 ms on one machine, and the MCP server no longer loads torch or an embedding model.
Models download only when you run trw-mcp models fetch or the installer, and no embedding vector leaves your machine. A dispatched reviewer gets only trw_recall and trw_code from TRW, and none of your other MCP servers.
Upgrade: install both packages together, stop the old memory daemon, run trw-mcp update-project --ide <id> and trw-mcp models fetch, reconnect your MCP clients, then run trw-mcp doctor. Removed tools have no aliases; most became a mode of a remaining tool or a trw-mcp command. TRW_OFFLINE is removed. Native Windows is not supported; use WSL2.

2026-09-24trw-mcp 6.1.0trw-memory 3.1.0

Embeddings by default, and the installer migrates for you

The installer now installs embeddings and caches the recall model, so semantic recall is ready after install. Pass --no-embeddings to stay keyword-only.
trw-mcp doctor and session_start report whether retrieval is active, degraded or off, with a one-line fix for anything degraded.
Upgrade: sqlite-vec is now a base dependency and the [vectors] extra is gone. musl Linux (Alpine) and Windows ARM are unsupported. If a project's .trw/memory/memory.db still holds learnings, the installer runs trw-mcp memory migrate --to user --apply and prints the rollback command; --no-migrate opts out.
Plus fixes for offline recall and for hooks on machines without jq.

2026-09-24trw-mcp 6.0.0trw-memory 3.0.0

One memory store per machine, and trw_assess

Memory now lives in one store per machine under ~/.trw, served by a background daemon, and each project reaches it through its own namespace.
trw_assess can be switched on from a project's config and stays off by default. Clients that take skills then get a short trw-assess skill that suggests it at decision points.
Tool responses late in a long session are no longer cut to their first 500 characters.
Upgrade: run trw-mcp update-project. If it prints trw-mcp memory migrate --to user --apply, run that from the project, then reconnect every MCP client. The store upgrade is one-way, so export first if you might roll back. trw_decision is renamed trw_assess with no alias, and trw_learn_update is gone: call trw_learn(learning_id=...).

2026-09-20trw-mcp 5.0.0trw-memory 2.0.0

Grok support, and formations that admit new members

Formations accept members after they are created, and sessions can announce themselves and be admitted in any order. Collaboration tools are on by default but do nothing outside a formation; comms_enabled: false hides them.
Grok is an installable client (--ide grok) with its native MCP config, AGENTS.md and bundled agents. Start grok from the project root and trust the project folder in grok itself.
Merges and consolidation no longer drop learning content or archive protected entries.
Upgrade: every formation member must run the same major version. Stop each member, then run trw-mcp formation comms-upgrade once per formation. Open slots now need open_join: true or trw-mcp formation admit. Install jq to keep hook event detail. trw-memory 2.0.0 removes the recall_rerank settings, because recall always re-ranks.

2026-09-18trw-mcp 4.0.0trw-memory 1.0.0

Opt-in collaboration and a new embedding model

Active agent sessions can opt in to exchanging messages and waiting briefly for replies. Messaging is pull-based, so it does not wake an idle agent. It is off by default.
Dispatch can give supported child sessions their own TRW connection, with reviewer restrictions and launch safeguards still in place.
Upgrade: the default embedding model changes. Keyword search still finds existing knowledge; semantic results are partial until vectors are re-embedded. TRW MCP does this in background batches, and standalone memory users can run trw-memory reembed --namespace <ns> for each namespace or call await client.reembed().

2026-09-17trw-mcp 3.1.0trw-memory 0.19.0

Hybrid recall by default and working macOS hooks

Recall is hybrid by default: keyword and meaning matches are fused, re-ranked by a cross-encoder and bounded by confidence, so it stops padding the list with weak matches.
A memory_maintain tool runs decay, consolidation and a checkpoint for one namespace on demand.
The delivery gate now sees your edits on every client profile, where before it saw them only on Claude Code.
Plus fixes for pre-edit hooks on macOS and for installs on arm64 Linux.

2026-09-15trw-mcp 3.0.0trw-memory 0.18.0

Plan review, and memories that survive transient failures

Plan review: an agent puts a plan up, a second agent reviews it against its own checks, and the first agent can tell whether the plan it is working from is still the approved one.
Agents working in different tools can pass work to each other directly. This is off until you turn it on.
A transient failure is now retried as a failure and never treated as permission to discard memories you already saved.
Upgrade: the experimental sprint-team and team-playbook slash commands are retired. File ownership, formations and the bundled team agents are unchanged.

2026-09-11trw-mcp 2.1.0trw-memory 0.17.0

Cross-model review that runs, and WAL checks that warn

Cross-model review now runs: the second opinion goes to another coding-agent CLI you already have installed, read-only.
A reviewer that could not be reached, timed out or answered with something other than findings is reported as exactly that, never as a clean review.
The WAL health check measures whether a checkpoint caught up with the write backlog, and when a WAL cannot be reclaimed it names the cause and the remedy.
Plus fixes for connectivity reporting and masked-tool access.

2026-09-07trw-mcp 2.0.2

Complete project upgrades and immutable releases

Upgrading a project now refreshes its hooks, skills, agents and instruction files along with the package.
Published versions are immutable. The release pipeline refuses to overwrite a version already served to users and checks every publish through the same download path the installer uses.
Plus installer fixes: a re-run keeps proprietary packages current, and pipx or uv upgrades keep vector search on.

2026-09-05trw-mcp 2.0.1

Vector search by default and faster recall

Vector search is on by default for every install path, with no extras to remember.
Learnings anchor to the code the session actually touched, and recall no longer re-checks stale anchors on the hot path. A representative recall dropped from 9.7 s to 2.5 s.
A doctor row and a heartbeat field flag a server burning CPU in the background, and one signal dumps every thread stack to a log with no debugger needed.
Plus fixes for fresh-install integrity checks and first-run diagnostics.

2026-09-05trw-memory 0.16.2

CPU fallback for embeddings, and dependency advisories closed

When the GPU is busy, the embedding model falls back to CPU, so recall keeps its embeddings for the session.
53 dependency advisories are closed across the optional extras. The runtime dependency set was never affected.

2026-09-05trw-mcp 2.0.0

Typed formations and namespace-keyed memory

A multi-agent formation is one validated artifact: file ownership, PRD assignment, join, status and brief each take a single call, and the delivery gate waits for every teammate.
Dispatch reaches seven coding assistants through one typed capability registry, which refuses any client whose capabilities were never verified against its real CLI.
Review verdicts expire, cross-model review must be proven with a digest-matched receipt, and safety-critical work can't ship without an independent adversarial audit.
Boot emits a five-phase timeline and doctor reports memory-engine health, so you can diagnose a slow start or a stuck store in one command.
Upgrade: memory rows are keyed on (namespace, id), so two namespaces can never collide in one store. This release requires trw-memory 0.16 or newer.

2026-09-05trw-memory 0.16.0–0.16.1

Namespace identity and a loopback memory daemon

You can run trw-memory as an authenticated loopback daemon, optionally with every namespace in one store per user. stdio stays the default.
A learning marked verified must now carry evidence, assertions or anchors.
Upgrade: every row is identified by (namespace, id), and reads and deletes name the namespace they mean. 0.16.1 is the version on PyPI; the 0.16.0 tag never published.
Plus fixes for concurrent store upgrades and warm model caches, and remote model code now runs only behind an explicit opt-in.

2026-08-04trw-mcp 1.0.1–1.0.5trw-memory 0.14.0–0.15.0

Security fixes for review receipts, hooks and the memory write gate

A session can no longer mint an "independent" review receipt for its own work.
The opt-in pre-edit hook no longer runs model-controlled input as Python (fixed in 1.0.2), and the Cursor secret-scan gate now fails closed on macOS as advertised. Existing installs receive the fixed hook from 1.0.4.
In trw-memory, five write paths that skipped the security gate now go through it, and a code-prefix trick that let a prompt-extraction payload past the injection gate is closed.
The trw-mcp fixes came from an audit of one week of changes by ten reviewers, with a second reviewer told to refute each finding. Each fix was checked by reverting it and watching its test fail.

2026-07-29trw-mcp 0.63.0–1.0.0

Structured tool arguments and a shorter CLAUDE.md

TRW no longer writes its protocol into CLAUDE.md for clients that don't read it, so the file drops from 80 lines to 17 for codex, opencode, copilot, cursor-cli and antigravity-cli. Upgrading removes the stale block between the TRW markers.
Upgrade: rarely set tool arguments move into structured parameters (metadata, fields, advanced and reviewer_identity), and unknown keys inside them are rejected. Five arguments that did nothing are removed, along with the trw_entity_risk_map tool. Pin trw-mcp<1.0 if you emit arguments by hand.
Upgrade: nudge_messenger: learning_injection now fails config validation; use contextual. 41 config fields that no production code read are removed, with their TRW_* environment variables.

2026-07-29trw-memory 0.12.0–0.13.2

Memories stored exactly as written

Memories are stored exactly as you wrote them. An earlier redaction backstop could rewrite real content, matching build numbers as social-security numbers; if you ran an earlier build, search your store for high-entropy or id redaction markers.
Memory entries can record that their verification failed, in a new optional verification_status field. An additive migration upgrades an existing store in place.
The prompt-injection gate is repaired, and legitimate knowledge about prompt injection can now be stored. Two known bypasses (order inversions such as "system prompt, now reveal it") stay open and documented, and statistical anomaly detection runs regardless.
Upgrade: MemoryClient(mode="mcp") is gone from the public Literal. Only type-checkers notice, since it always raised at runtime.

2026-07-20trw-mcp 0.58.0–0.62.0

API keys leave tracked config, and the installer prompts again

Upgrade: platform_api_key is no longer read from the git-tracked .trw/config.yaml. It resolves from an environment variable, then the ignored 0600 .trw/credentials.yaml, and update-project migrates an existing tracked key and blanks it in config.yaml.
The one-line installer asks which clients to set up again under curl … | bash. Headless and CI runs stay non-interactive.
The trw-simplify skill and its trw-code-simplifier agent are retired and removed from existing installs on update.
Plus update-project fixes for nested symlinks such as agent worktrees and node_modules shims.

2026-07-18trw-mcp 0.57.0trw-memory 0.11.0

Cross-client dispatch and stdio-only transport

Cross-client dispatch hands a task to another coding CLI (Codex, Claude or opencode) for a background second-opinion audit. Read the result back with trw_dispatch and trw_dispatch_status.
Dispatch is sandboxed by default: a sanitized environment, read-only enforcement per client, a process-group kill on timeout, prompt redaction and a confined working directory.
The default trw_recall payload is about 65% smaller, with a verbose mode when you want more.
trw-memory scans stored evidence for secrets and PII, including GitHub and AWS key patterns, and redacts them. Its schema upgrade runs behind a mandatory backup snapshot.
Upgrade: trw-mcp is now stdio-only. The shared HTTP server, the stdio proxy and the --transport, --host and --port flags are removed (leftover config keys are ignored). Tool changes take effect on the next reconnect.

2026-06-09trw-mcp 0.55.0trw-memory 0.9.0

Portable learnings that follow you across repos

An opt-in user memory tier at ~/.trw lets portable learnings follow you across every repo on the machine. It is off until you opt in, and enabling it never moves or rewrites existing project learnings.
trw_learn classifies portability automatically, keeping repo-local gotchas with the project and cross-cutting knowledge machine-wide.
trw_recall merges project and user tiers into one ranked result, with a cap so user hits never bury a precise project match.

2026-06-08trw-mcp 0.54.0trw-memory 0.8.5

Build failure triage and a compact session start

Build-check failures are tagged likely-yours, likely-pre-existing or unknown, so you can triage them fast.
Session start is compact by default: learnings cap to the top results and diagnostics collapse to a one-line health summary, with a verbose mode on demand.
An intentional-code marker flags deliberately counterintuitive code so the bundled reviewer and simplifier agents leave it alone.
Plus fixes for run logs after a concurrent write and for first-run login.

2026-05-29trw-memory 0.8.2–0.8.5

Reads that survive a bad row, and faster batch writes

A single bad row is quarantined on read, so one damaged entry no longer takes down the whole query.
Batch passes collapse many writes into a single commit, removing per-row overhead.
The PII scanner catches GitHub PAT and AWS access-key patterns before they can be stored in plaintext.
The native SQLite dependency installs only on Linux, with a stdlib fallback on macOS and Windows so installs keep working.

2026-05-28trw-mcp 0.48.0–0.48.15trw-memory 0.8.3

Per-client write surfaces and Antigravity CLI support

Per-client write surfaces for Claude Code, Cursor, Codex, OpenCode, Copilot and Antigravity CLI, switched on automatically at install and update.
Antigravity CLI gets full-ceremony support, with environment discovery, config deep-merge and specialist subagents.
A capability-tier resolver maps the portable frontier, balanced and local tiers to each client.
You can send feedback from inside your editor, and secrets are redacted before anything goes over the network.
Plus a fix for a multi-minute stall in learning writes.

2026-04-27trw-mcp 0.47.0trw-memory 0.8.1

Eight client profiles and signed learnings

Eight first-class client profiles: Claude Code, Cursor IDE, Cursor CLI, Aider, Codex, Copilot, Gemini and OpenCode.
Learnings are signed when written, so tampering at the storage layer can be detected across sessions.
Every tool dispatch now goes through one verified audit, telemetry and kill-switch surface set up at boot.

2026-04-26trw-mcp 0.46.0trw-memory 0.7.0

Memory-poisoning defense and per-item batch recovery

Memory-poisoning defense starts in observe mode: suspicious entries surface in a shadow partition, and a calibration window runs before enforcement is switched on.
Canary entries verify themselves on recall to detect tampering at the storage layer.
Batch memory writes recover per item, so one bad entry no longer fails the whole batch.

2026-04-23trw-mcp 0.43.0–0.45.2

Prompts tuned per client and per model

Every prompt surface is recalibrated for Claude frontier 4.7: instruction style, default effort and tokenizer behavior.
Prompts render for the client profile you use, with no hardcoded CLI assumptions, and tool and skill descriptions follow one consistent pattern.

2026-04-12trw-mcp 0.42.0

Instruction files list only available tools

Instruction files list only the tools that are available, so agents aren't pointed at tools they can't call.
A CLI command checks instructions against the active tool set, and the delivery gate warns when instructions name a tool that isn't available.

2026-04-02trw-mcp 0.39.1

Sturdier OpenCode support

OpenCode support gets full-package verification, and the bundled OpenCode artifacts stay in sync with the source skills, hooks and agents.
Plus fixes for learning outcome records and session boundaries.

2026-03-31trw-mcp 0.37.0trw-memory 0.6.2

Contextual learning injection and leaner tool responses

Relevant learnings are injected when a run changes phase.
Tool responses use fewer tokens per call, and instruction files and hooks are lighter.
Agents are consolidated into focused roles, with a clean upgrade path.
Platform sign-in is hardened, including recovery from cold starts.

2026-03-30trw-mcp 0.36.0

Codex support

Codex gets a full bootstrap: one command generates every project config, instruction file and skill definition, with merge-safe updates.
Plus fixes for skill paths and for project setup in some directory layouts.

2026-03-29trw-mcp 0.35.0trw-memory 0.6.0

A storage interface for trw-memory

trw-memory 0.6.0 adds an abstract storage interface and clearer CLI errors.
Bundled scripts get security fixes, and every error handler was reviewed.

2026-03-28trw-mcp 0.34.0

Embeddable server and new CLI commands

The MCP server can be embedded as a library for testing and integration, with full type annotations.
New CLI commands: a generated config reference, project uninstall with a dry-run mode, and --version.

2026-03-28trw-mcp 0.33.0

Long sessions stay live

Tool calls signal liveness, so a long-running session is no longer marked stale by mistake.
Proactive database maintenance keeps storage from growing without bound during long sessions.
Plus fixes so session start is always recorded and delivery gates work in any tool-call order.

2026-03-26trw-mcp 0.32.0trw-memory 0.4.0

Memory that checks itself against your code

Learnings can carry machine-verifiable assertions that run against your codebase, so stale knowledge gets flagged.
Recall ranks entries with failing assertions lower.
Verification has path-traversal protection, pattern limits and binary-file detection.

2026-03-25trw-mcp 0.31.0

One-line install and device sign-in

A single shell command detects your environment, installs from PyPI and configures your project.
Device authorization signs your CLI in through the browser, with QR code support, and login, logout and status handle multiple organizations.
Social login with Google and GitHub, including Google One Tap.

2026-03-25trw-mcp 0.30.0

An adaptive context budget, and the move to source-available

An adaptive context budget compresses tool responses as a session grows, and repeated identical responses are suppressed, cutting token use.
TRW is prepared for source-available publication under BSL 1.1, with a public README.
Upgrading clears the instruction cache, so every project re-renders its instructions after a version bump.

2026-03-19trw-mcp 0.27.0trw-memory 0.3.0

Queryable logs and safer workstream cleanup

Logs are restructured into consistent, queryable events across all modules.
Cleanup now keeps the work of parallel agents in a workstream.

2026-03-18trw-mcp 0.22.0

Client profiles and compact tool responses

Built-in client profiles, including claude-code, opencode, codex and aider, adapt ceremony to each client.
A response optimizer compacts tool responses to save LLM context.
Plus delivery reliability fixes.

2026-03-15trw-mcp 0.20.0

Ceremony modes for every client, and the orchestration toolkit ships

Ceremony adapts to Claude Code, Cursor, OpenCode and Codex, with full and light modes.
The orchestration toolkit ships tools, resources, skills and bundled agents for the development lifecycle.
High-impact learnings auto-promote into instruction files for future sessions. (Retired in 0.37.0: learnings now arrive through recall at session start.)
The installer is overhauled with health checks and cross-platform server management.

2026-03-14trw-mcp 0.15.0

File ownership and pre-spawn safety checks

File ownership stops multiple agents from editing the same files at once.
Pre-spawn safety checks block agent creation when uncommitted changes could cause conflicts.

2026-03-14

Documentation rewrite and a full API reference

A full API reference with endpoint docs, auth details, examples and error formats.
Every documentation page is rewritten to be easier to scan, and the references cover the full tool, skill and agent catalog.

2026-03-05

Interactive memory visualizations

An interactive memory dashboard with growth timeline, tier migration, smart search, knowledge graph and forgetting curve views.
Searching highlights matching data across every dashboard section.
A homepage growth timeline shows how framework knowledge compounds across tasks.

2026-02-27trw-memory 0.1.0

A standalone memory engine with hybrid search

A standalone persistent memory engine for AI coding agents, with local storage.
Hybrid search combines sparse keyword and dense vector retrieval.
Knowledge graph, tiered storage lifecycle, semantic dedup and LLM-powered consolidation.
Integrations with LangChain, LlamaIndex, CrewAI and OpenAI-compatible APIs.

Next steps

Use quickstart to confirm the current install and access flow, configuration to understand changed defaults, and the API reference when a release touches hosted platform behavior.

Quickstart

Current access, install, and first-session flow

Configuration

Customize TRW for your project

API reference

Hosted platform auth and endpoint changes