Memory with layers.
Truth on top.

Governed runtime memory for coding agents — supersession keeps current truth on top, scopes isolate context, audit export shows what changed. Local SQLite, MCP stdio, spec v1.7.0 with 143 validation behaviors.

Documentation Read the spec

New facts supersede old ones — history is kept, but recall returns only what is current.

What goes wrong with append-only memory

Common failure modes in long-running agent sessions.

Stale facts

The agent keeps citing an old stack or a decision you reversed. Palimem records a new value and supersedes the old one instead of leaving both in the retrieval set.

Scope bleed

User notes, session scratch, and repo facts use separate scopes so search does not mix contexts.

No audit trail

Point-in-time recall and audit export show what was current when, not only what ranks first today.

Different problem, different architecture

Palimem optimizes governed correct state, not recall@k leaderboard scores.

Cloud memory APIs Hosted retrieval layer; data leaves your machine

Local-first governed memory SQLite on your disk — no mandatory cloud account

Append-only MCP dumps Facts pile up; ranking picks a winner at read time

Supersession on write Current truth on top; history preserved for audit

Benchmark chasers Optimize recall@k on public leaderboards

Correct state per subject One current value per key — scoped, auditable, spec-tested

Core behavior

Specified in spec v1.7.0 and covered by the validation suite.

Correct by design

Supersession on write, structured subject keys, and dependent-field propagation — updates stay consistent without a nightly consolidation gamble.

Scoped recall

User, session, and repository scopes keep context where it belongs. Optional subject filters reduce same-field pollution.

Connect anywhere

npm package, Claude Code plugin, or MCP stdio — Palimem meets you where you code. Examples for Claude Code, Cursor, Copilot, and Codex. Adapters for Hermes and OpenClaw.

Audit-ready

Temporal queries, as_of recall, PII scan, retention controls, and audit export — for teams that need to show their work.

Connect your harness

Palimem meets you where you code — plugin, MCP, or adapter. Example configs in the repo; no cloud account required.

MCP stdio

Claude Code

Recommended path. Copy examples/claude-code/.mcp.json to your project root — smoke-tested in CI.

View setup guide

Spec and tests

Open source under Apache-2.0. v1.7.2 release · @palimem/mcp · component spec

v1.7.0

Published spec

143/143

Validation behaviors

MCP tools in spec

Apache-2.0

Open-source license

Frequently asked questions

How does Palimem relate to OKF or AGENTS.md?

They compose — Palimem does not replace them. OKF bundles and AGENTS.md carry curated project knowledge (what the project is). Palimem governs runtime memory (what the agent should believe now across sessions, with supersession and audit). See Palimem and OKF and why governed memory.

Is Palimem a vector database?

No. Palimem stores governed records and rebuildable lexical indexes. Embeddings are optional; the default path does not require a vector database.

Do I need a cloud account?

No. Data stays in SQLite on disk under your project or home directory. Connect via MCP stdio or the bundled adapters.

Do I need MCP?

No. MCP is the easiest path for Claude Code and Cursor. Palimem also runs as a Python service with adapters for Hermes and OpenClaw. A REST API for platform integrations is on the roadmap.

How is this different from Mem0 or Zep?

Palimem is local-first governed memory: supersession on write, scoped recall, and an audit trail in SQLite on your disk. Mem0 and Zep are often cloud-hosted retrieval layers; Palimem does not require a cloud account and optimizes for correct current state, not recall@k leaderboard scores.

What harnesses are supported?

Spec v1.7.0 documents eleven integrations with example configs and smoke tests: Claude Code, Claude Code plugin, Copilot CLI, Codex, Cursor, Windsurf, VS Code + Copilot Agent, Copilot IDE, Gemini CLI, Hermes, and OpenClaw. See docs/getting-started and per-harness guides.

Is it production-ready?

Spec v1.7.0 ships with 143 automated validation behaviors and integration smoke tests for every tier B+ harness in the catalog. Install from npm (npx @palimem/mcp) or from source on GitHub.

How do I install from npm?

@palimem/mcp v1.7.2 is on npm. Connect an editor without cloning:

npx @palimem/mcp ai-memory connect <harness> \
  --project-root "$(pwd)" \
  --launcher npx \
  --data-dir .ai-memory/data

See the installation guide. PyPI packages are not yet published. Fallback: npx github:palimem/palimem.

Get started

Apache-2.0. Install from npm (@palimem/mcp) or GitHub. Local SQLite, MCP stdio.

Documentation View on npm → View on GitHub →

Memory with layers.Truth on top.