cogmem
A self-improving, verifiable memory layer for AI coding agents.
cogmem learns how you work across sessions so your agent gets more accurate and more autonomous over time: it stops repeating mistakes, keeps a live model of each project, and surfaces the right lesson at the right moment. Every memory is cryptographically signed and tamper-evident, so a poisoned or altered memory can be detected and rejected before it ever steers the agent.
Developed by WritersLogic — local-first recall; your memory and identity key stay on your machine (see Privacy).
Installation
From PyPI
pip install cogmem
pip install 'cogmem[recall]'
cogmem init
pip install cogmem gives you the cogmem CLI and the MCP server (cogmem mcp, or uvx cogmem mcp on demand) — the verifiable-memory tools need only the core install. Add the [recall] extra for local semantic recall, then run cogmem init to wire the full learning loop (the SessionStart/UserPromptSubmit/Stop hooks and the index) into Claude Code. cogmem init is idempotent; re-run it any time.
Clone installer (turnkey, with the warm daemon)
The clone installer does everything pip install + cogmem init does, plus sets up the warm recall daemon (launchd/systemd) as a managed service:
git clone https://github.com/writerslogic/cogmem.git
cd cogmem
./install.sh
Or in one line:
curl -fsSL https://raw.githubusercontent.com/writerslogic/cogmem/main/install.sh | bash
install.sh is idempotent — run it again any time to upgrade in place. It sets up
the code under ~/.claude/cogmem, a self-contained virtualenv with dependencies,
the cogmem CLI on your PATH, the Claude Code hooks, and a warm recall daemon
(a launchd agent on macOS, a systemd --user service on Linux). Requires
Python 3.12+; semantic recall runs on a local model (fastembed, no external
API). Pass --no-daemon or --no-hooks to skip those steps; set COGMEM_HOME
to install elsewhere — the CLI, engine, and hooks all resolve it at runtime, so
a non-default install keeps its memory and identity fully self-contained.
Custom Installation
Install to a different directory
Set COGMEM_HOME to place cogmem somewhere other than the default
~/.claude/cogmem:
COGMEM_HOME=/opt/cogmem ./install.sh
Or with the one-liner:
curl -fsSL https://raw.githubusercontent.com/writerslogic/cogmem/main/install.sh | COGMEM_HOME=/opt/cogmem bash
The installer copies the code, creates the virtualenv, and symlinks the CLI to
~/.local/bin/cogmem (or wherever COGMEM_BIN points).
CLI path
Set COGMEM_BIN to control where the cogmem CLI symlink is placed:
COGMEM_BIN=$HOME/.cargo/bin COGMEM_HOME=/opt/cogmem ./install.sh
If COGMEM_BIN is not on your PATH, the installer prints a warning. You can
always invoke cogmem directly from $COGMEM_HOME/cogmem.
How data directories and identity keys are resolved
At runtime the CLI and engine read COGMEM_HOME from the environment. When it is
unset they fall back to ~/.claude/cogmem. All runtime data lives under the
vault/ subdirectory:
| Path | Purpose |
|---|
$COGMEM_HOME/vault/identity/agent.key | Ed25519 private key (agent identity, did:key) |
$COGMEM_HOME/vault/credentials/ | W3C Verifiable Credential storage |
$COGMEM_HOME/vault/rules/ | Layer-A (always-load) and Layer-B (recall) rules |
$COGMEM_HOME/vault/provenance/log.jsonl | Append-only hash-chained transparency log |
$COGMEM_HOME/vault/provenance/statements/ | COSE_Sign1 SCITT signed statements |
$COGMEM_HOME/engine/.venv/ | Python virtualenv with dependencies |
$COGMEM_HOME/hooks/ | Claude Code hook scripts |
The identity key is generated on first run (via cogmem status or any engine
operation) and persisted at $COGMEM_HOME/vault/identity/agent.key. The
corresponding did:key is derived from the Ed25519 public key. Moving or
reinstalling cogmem to a new COGMEM_HOME creates a fresh identity unless you
migrate the vault/ directory.
MCP client with a non-default install
The standard MCP client configuration works regardless of COGMEM_HOME because
the cogmem CLI resolves the environment variable at runtime:
{
"mcpServers": {
"cogmem": { "command": "cogmem", "args": ["mcp"] }
}
}
If the CLI is not on your PATH, use the full path:
{
"mcpServers": {
"cogmem": { "command": "/opt/cogmem/cogmem", "args": ["mcp"] }
}
}
Or prefix with COGMEM_HOME in a shell wrapper:
{
"mcpServers": {
"cogmem": { "command": "env", "args": ["COGMEM_HOME=/opt/cogmem", "cogmem", "mcp"] }
}
}
Quick Start
cogmem status
cogmem doctor
cogmem recall "..."
cogmem note "..."
cogmem verify
cogmem receipt <id>
cogmem statement <id>
cogmem trust
cogmem trust --rotate
cogmem review list
cogmem mcp
MCP Integration
Run cogmem as an MCP server and connect any MCP-compatible client:
{
"mcpServers": {
"cogmem": { "command": "cogmem", "args": ["mcp"] }
}
}
Eight tools are exposed: recall, note, status, verify, receipt, tree_head, progress, review_pending, plus read-only resources (the live user model and per-project state).
Claude Code Integration
install.sh wires cogmem into Claude Code automatically (idempotently merged into
~/.claude/settings.json) — no manual invocation required. Five hooks make the
memory loop run in the background:
| Event | Hook | What it does |
|---|
SessionStart | cogmem-activate.sh | injects promoted always-load (Layer-A) rules + the self-check |
UserPromptSubmit | cogmem-recall.sh | semantic Layer-B recall for the current prompt |
PreToolUse(Bash) | cogmem-guard.sh | intercepts known mistakes at the tool-call boundary before they happen |
PostToolUse(Edit|Write) | cogmem-context.sh | tracks which files the session is actively editing |
Stop | cogmem-capture.sh | captures the session into memory (acquisition + consolidation) |
Every hook is strictly fail-open: any error, timeout, or cold daemon injects
nothing and never blocks your prompt. The scripts live in ~/.claude/cogmem/hooks/;
re-run install.sh (or ./install.sh --no-daemon) to refresh the wiring.
Why cogmem? -- learns from outcomes, models failure modes, verifiable memory
Chat-memory systems (Mem0, Letta, Zep) store and retrieve facts. cogmem is built for coding agents and goes further on three axes:
It learns from outcomes. A feedback loop scores whether a recalled lesson actually helped, refines rules that prove wrong, and retires ones that mislead.
It models its own failure modes. cogmem tracks where the agent tends to go wrong in your work and intercepts known mistakes at the tool-call boundary — before they happen, not afterward.
Its memory is verifiable. Each memory is a W3C Verifiable Credential signed by the agent's did:key, recorded in a tamper-evident, SCITT-style transparency log. Agent memory is an attack surface; cogmem makes it auditable and poison-resistant.
Features -- two-layer memory, outcome feedback, self-model, project state, cross-project narrative, self-regulation, verifiable credentials
- Two-layer memory: always-loaded directives (scope-gated, human-approved) plus a semantic recall tail (local cross-encoder reranking, no data leaves the machine).
- Outcome feedback and self-refinement: memories earn or lose trust based on whether they actually helped; contradicted rules are corrected through a safe pipeline.
- Self-model and guard: a model of the agent's recurring mistakes, compiled into tripwires that intercept them at the
PreToolUse boundary.
- Project-state model: a living per-project state (goal, claims, open questions, blockers) that gives situational continuity and reasons across time.
- Cross-project progress narrative: momentum, stalls, and dependencies across projects, surfaced as alerts.
- Self-regulation: recall thresholds tuned automatically against an eval harness.
- Verifiable Agent Memory:
did:key identity, W3C VC-signed memories, COSE_Sign1 SCITT signed statements (byte-compatible with HMS), a hash-chained transparency log with signed Merkle tree head and RFC 6962 inclusion receipts, optional poison-resistance enforcement. See PROVENANCE.md.
Verifiable Memory -- did:key identity, W3C VC, COSE/SCITT, hash-chained log, poison-resistance
cogmem treats every stored memory as a signed artifact:
did:key identity: each agent gets a persistent Ed25519 identity, exposed as a W3C DID.- W3C Verifiable Credentials: every memory is signed with
eddsa-jcs-2022 Data Integrity proofs.
- COSE_Sign1 / SCITT signed statements: byte-identical to the envelope format used by holographic-memory and crosstalk — independently verifiable by any of the three implementations.
- Hash-chained transparency log: append-only JSONL with SHA-256 chaining, a signed Merkle tree head, and RFC 6962-style inclusion receipts.
- Poison-resistance: altered or injected memories fail verification and are rejected before influencing the agent.
cogmem verify
cogmem receipt <memory-id>
See PROVENANCE.md for the full specification.
Verify the C2PA sample yourself:
./examples/c2pa-agent-credential/verify.sh
This proves the whole chain: agent identity (cawg.ica.credential_valid) bound to real cognition — a signed cogmem memory and a signed crosstalk reasoning audit, each an independently verifiable Ed25519 COSE/SCITT statement.
Privacy
cogmem is local-first by design. Memories, embeddings, and the identity key live on your machine, and semantic recall is fully local — the embedding and reranker models (fastembed) run on-device, so querying your memory never leaves the machine.
The learning pipeline is not local: acquisition, consolidation, the feedback judge, and the project/user-model synthesis send the relevant session transcript to the Anthropic API (ANTHROPIC_API_KEY). That is how rules are extracted and scored. If you need fully-offline operation, run with --no-hooks (recall still works) until a local-model extraction path lands. In short: recall is local; learning calls the API.
Part of the Agent-Provenance Stack
cogmem is one component of the WritersLogic verifiable agent-provenance pipeline — agent identity, memory, reasoning, and signed output, cryptographically bound end to end.
| Project | Role |
|---|
| cogmem (this repo) | Agent identity (CAWG credential) + verifiable, tamper-evident memory (COSE/SCITT) |
| crosstalk | Multi-model orchestrator; signs each turn's reasoning/orchestration audit |
| holographic-memory | Durable holographic memory store; cross-verifies signed statements and agent identity |
| WritersProof | C2PA producer: binds identity + memory + reasoning to the signed asset |
All four share one substrate — COSE_Sign1 / SCITT signed statements (Ed25519) and W3C DID identity — specified in UNIFIED-PROVENANCE.md.
License
Apache-2.0 — see LICENSE.