agenr
v2026.6.5
Published
Agent memory - local-first knowledge infrastructure for AI agents
Readme
█████╗ ██████╗ ███████╗███╗ ██╗██████╗
██╔══██╗██╔════╝ ██╔════╝████╗ ██║██╔══██╗
███████║██║ ███╗█████╗ ██╔██╗ ██║██████╔╝
██╔══██║██║ ██║██╔══╝ ██║╚██╗██║██╔══██╗
██║ ██║╚██████╔╝███████╗██║ ╚████║██║ ██║
╚═╝ ╚═╝ ╚═════╝ ╚══════╝╚═╝ ╚═══╝╚═╝ ╚═╝
AGENt memoRyagenr
Local-first memory for AI agents. One SQLite brain that survives restarts, tools, and sessions - shared across OpenClaw, Skeln, and the CLI.
Most runtimes forget what mattered yesterday. agenr keeps memory structured, searchable, and on your machine. Only model and embedding calls leave the box.
Memory at a glance
agenr splits agent memory into layers. Each answers a different question.
| Layer | Question | In short | | -------------- | ---------------------------- | -------------------------------------------------------------------------------------------------------- | | Durable | What is true? | Facts, decisions, preferences, lessons - distilled knowledge with claim-key lifecycle and hybrid recall. | | Episodic | What happened? | Session-level narrative summaries for time-bounded questions like "what did we do last week?" | | Working | What are we doing right now? | Transient task state - scratchpad, checkpoints, next actions - injected per turn, not durable truth. | | Procedural | How do I do this? | Repo-authored YAML runbooks synced into the store for repeatable workflows. | | Dreaming | Is the corpus healthy? | Background maintenance: scan, reconcile, and repair durable memory behind an explicit apply gate. |
Durable and episodic live in the database. Working is session-scoped and fades when the task moves on. Procedural is authored in git and synced in. Dreaming keeps the long-term store coherent over time.
Deep dives:
- Durable memory - store pipeline, claim keys, supersession
- Episodes - session summaries and temporal recall
- Working memory -
agenr_work, checkpoints, transient context (Skeln details) - Procedures - authoring, sync, and recall routing
- Dreaming - tiers, scan/reconcile/apply, corpus health
- Recall - hybrid search across memory layers
Quick start
pnpm install -g agenr
agenr initThe wizard sets up auth, embeddings, your database (~/.agenr/knowledge.db), optional OpenClaw plugin install, and an initial transcript ingest pass. Run it again any time to reconfigure or ingest more sessions.
Manual plugin install
# OpenClaw
openclaw plugins install @agenr/openclaw-plugin
openclaw gateway restart
# Skeln - see docs/SKELN-PLUGIN.md for packaging and config
skeln extension add @agenr/skeln-pluginTry recall from the CLI
agenr recall "what decisions did we make about the API?"
agenr ingest ~/.openclaw/agents/main/sessions/ # durable extraction
agenr ingest episodes --recent 30d # episodic backfill
agenr dream status # corpus healthFull CLI reference: AGENTS.md.
Host integrations
| Host | Package | Agent tools |
| -------- | ------------------------ | -------------------------------------------------------------------------- |
| OpenClaw | @agenr/openclaw-plugin | agenr_store, agenr_recall, agenr_fetch, agenr_update, agenr_work |
| Skeln | @agenr/skeln-plugin | above + get_goal, create_goal, update_goal |
Both plugins share the same database and recall brain. Details: OpenClaw plugin, Skeln plugin.
Documentation
| Topic | Doc |
| ------------------------------- | ---------------------------------------------- |
| Architecture and repo shape | docs/ARCHITECTURE.md |
| Ingest pipelines | docs/INGEST.md |
| Configuration and env overrides | agenr setup or agenr init |
| Contributing / development | AGENTS.md |
| Debugging | docs/DEBUGGING.md |
Config lives at ~/.agenr/config.json by default. Override with AGENR_CONFIG_PATH, AGENR_CONFIG_DIR, or AGENR_DB_PATH.
Development
pnpm install
pnpm build
pnpm check # format, lint, typecheck, testSandbox helpers and the full contributor workflow are in AGENTS.md.
License
MIT
