Repository navigation
Releases
Mnemo 0.1.5
A security fix for the installer.
Fixed
mnemo installno longer prints other tools' credentials. Its preview shows each config change as a diff, and the lines around a change (and OpenCode's config, which it rewrites whole) could include tokens and API keys that other tools keep in the same files, for example a GitHub token in~/.codex/config.toml. Those values now show as<hidden>: settings whose names look like credentials, and values shaped like known tokens (github_pat_,ghp_,sk-,xoxb-,hf_and similar). The files themselves keep their values.
If you ran mnemo install with 0.1.4 or earlier in a shared terminal or recorded session, check its output for credentials and rotate any that appear.
Upgrading
uv tool upgrade mnemohiveNo new migrations and no other changes since 0.1.4.
Mnemo 0.1.4
Mnemo is now on PyPI as mnemohive. The command is still mnemo. This is the first PyPI release; it carries everything listed for 0.1.3, which was released on GitHub only, plus the fixes below.
Installing
uv tool install mnemohive
mnemo installInstalled from GitHub before 0.1.3? Remove the old tool first, because it used the name mnemo:
uv tool uninstall mnemo
uv tool install mnemohive
mnemo install # applies the new migrations to the memories you already haveFixed since 0.1.3
- A memory can't be replaced by its identity through SQL. An
INSERT OR REPLACEwith a new ID and an existing memory's scope and key replaced that memory (mnemo verifyreported it afterwards). Migration0004_fact_identity_unique.sqlnow refuses it. - The near-duplicate warning reports only live memories, not ones that are not yet valid, ended or expired, and shows numbers and JSON values as text.
mnemo installchecks SQLite's math functions, which search ranking needs, instead of search failing later.- A bare
/fact/<id>link in the web UI opens a project memory in its own scope.
What's new
Defence against wrong or poisoned memories
- A memory's identity is permanent. SQL can no longer move a memory to another project, rename it or delete it, and each memory's identity is sealed in a second hash chain.
mnemo verifyreports one changed around the triggers. mnemo verifyremembers the head it verified. It stores the head beside the store (mnemo.db.verified, or--witness FILE), so a rewrite of the whole history is caught next time. After restoring a backup on purpose, delete that file.- The memory policy and the server's instructions treat files, command output and memories as data. Agents report planted instructions to you, and retract a memory that carries instructions for assistants. Measured on four poisoning scenarios (eval).
- Corrections in the web UI are guarded. If an agent changed the memory while the page was open, your correction is refused instead of overwriting it. The UI also refuses cross-site form posts.
Better memory
- Search matches any word of a question, ranked by how many rare words a memory shares. More memories now come up as the first result: 21 of 28 questions on both models, from 20 and 19.
- Near-duplicate warning. Saving a memory that says nearly the same as an existing one in the same project returns the existing ones, so the agent can keep a single memory.
- Memories archived by the old decay job can be restored, from the web UI's Retracted page or with update or revert.
Everyday use
- The web UI has a Scope menu for the global memories and each project's.
- The local embedding model is released after 5 idle minutes, which frees about 90 MB per idle agent session on Linux.
- Linux is supported and checked:
scripts/check-linux.shruns a full install in Docker. Windows has been prepared but not tested. mnemo verifyandmnemo installwarn when the store's directory is readable by other users.
Agent eval
With all of this, the 16-scenario eval with Claude Code and Codex passes 16 of 16 (results).
Mnemo 0.1.3
Superseded by 0.1.4, the first release on PyPI. 0.1.3 was never uploaded there; install 0.1.4 with uv tool install mnemohive. The changes below are all in 0.1.4.
What's new
Defence against wrong or poisoned memories
- A memory's identity is permanent. SQL can no longer move a memory to another project, rename it or delete it, and each memory's identity is sealed in a second hash chain.
mnemo verifyreports one changed around the triggers. mnemo verifyremembers the head it verified. It stores the head beside the store (mnemo.db.verified, or--witness FILE), so a rewrite of the whole history is caught next time. After restoring a backup on purpose, delete that file.- The memory policy and the server's instructions treat files, command output and memories as data. Agents report planted instructions to you, and retract a memory that carries instructions for assistants. Measured on four poisoning scenarios (eval).
- Corrections in the web UI are guarded. If an agent changed the memory while the page was open, your correction is refused instead of overwriting it. The UI also refuses cross-site form posts.
Better memory
- Search matches any word of a question, ranked by how many rare words a memory shares. More memories now come up as the first result: 21 of 28 questions on both models, from 20 and 19.
- Near-duplicate warning. Saving a memory that says nearly the same as an existing one in the same project returns the existing ones, so the agent can keep a single memory.
- Memories archived by the old decay job can be restored, from the web UI's Retracted page or with update or revert.
Everyday use
- The web UI has a Scope menu for the global memories and each project's.
- The local embedding model is released after 5 idle minutes, which frees about 90 MB per idle agent session on Linux.
- Linux is supported and checked:
scripts/check-linux.shruns a full install in Docker. Windows has been prepared but not tested. mnemo verifyandmnemo installwarn when the store's directory is readable by other users.
Agent eval
With all of this, the 16-scenario eval with Claude Code and Codex passes 16 of 16 (results).
Mnemo 0.1.2
Two additions to Mnemo's defence against wrong or poisoned memories.
What's new
- Retracted memories in the web UI.
mnemo uihas a Retracted page listing every memory an agent withdrew withmemory_retract: its value, who retracted it, when and why. Open one to restore an earlier revision, so a retraction an agent was misled into is visible and reversible. - Hash-chained history and
mnemo verify. Every event is now linked to the previous one by a SHA-256 hash.mnemo verifyrecomputes the chain and reports any history changed, removed or inserted outside Mnemo (for example by editing the file with thesqlite3CLI), plus any memory whose current value is not its newest revision. It prints the chain head: keep it, and a later rewrite of the whole chain shows up too. - Searching in the web UI no longer counts as an agent recalling a memory.
Upgrading
uv tool install --force git+https://github.com/Yugesh-reddy/Mnemo
mnemo install # applies the new migration and chains the history you already haveUntil the migration is applied, the agents' memory tools return an error saying to run mnemo install. Nothing else changes for the agents: the seven tools behave as before.
Limit
The hash chain detects edits made around Mnemo; it cannot stop someone who rewrites the file and recomputes every later hash. Comparing with a chain head you kept catches that.
Mnemo 0.1.1
Mnemo is one shared, versioned memory for the coding agents on your machine: Claude Code, Codex and OpenCode read and write the same store through seven MCP tools, and every change is attributed, kept in history and reversible.
0.1.x replaces Postgres with one local file. No Docker, no database server, no password, and by default no Ollama: memory is ~/.mnemo/mnemo.db (readable only by you), and the embedding model runs inside Mnemo.
uv tool install git+https://github.com/Yugesh-reddy/Mnemo
mnemo install # shows every change, asks firstWhat's new in 0.1.0 and 0.1.1
- One SQLite file instead of Postgres + pgvector. History, guarded updates, revert, retract, sources and retry receipts behave as before. Every write takes the file's single write lock: eight processes racing one update give exactly one winner.
- The embedding model runs inside Mnemo (
BAAI/bge-small-en-v1.5, a 67 MB one-time download).--backend ollamakeeps using Ollama. - Model switches keep your memories. Each vector records the model that made it;
mnemo installembeds stored memories again after a switch, andmnemo reindexdoes it by hand. - New commands:
mnemo export,mnemo import,mnemo reindex.mnemo upandmnemo downare gone. - Runs on any Python 3.12+ (0.1.1): where SQLite cannot load extensions, as in the python.org builds, Mnemo computes the same cosine scores in Python.
- CI on macOS and Linux.
Upgrading from the Postgres version
The new version does not read the Postgres database; carry your memories over with an export, which keeps every ID, revision and retry receipt:
mnemo-transfer export --all-scopes --out mnemo-postgres.json # old version, Postgres running
uv tool install --force git+https://github.com/Yugesh-reddy/Mnemo
mnemo import mnemo-postgres.json
mnemo install # add --backend ollama to keep using OllamaYour Postgres data is not changed. To go back: uv tool install --force git+https://github.com/Yugesh-reddy/Mnemo@v0.0.1-postgres.
Evidence
- The coding-agent eval (real Claude Code and Codex sessions, 12 scenarios including two memory-poisoning attempts) passed 12 of 12 in three runs against the 0.1.0 wheel, with no Mnemo tool errors. Write-up
- Search over 5,000 memories in a project takes about 31 ms (about 48 ms on the Python fallback).
Known limits
- The default model returns a memory for an unrelated question more often than Ollama (3 of 8 such questions on the retrieval set, against 1). Agents see each hit's source and trust.
- Each agent session's server uses about 300 MB once the model loads.
- Tested on macOS; Linux passes CI but a full agent install there is untested. Windows is untested.
- Local, single-user: memories are stored unencrypted, so don't save secrets.