Persistent, self-evolving AI memory plugin for coding agents and applications. Combines LLM-driven memory extraction with a two-layer Experience + Skill distillation system: raw interactions are first compressed into Experience memories, then recurring patterns are further abstracted into reusable Skill entries. Ebbinghaus-style time-decay keeps memory collections pruned and relevant over time. Exposes a unified backend via Python SDK, HTTP REST server, MCP server, and CLI.
- Storage
- Multi-backend storage: OceanBase (native), PostgreSQL, and SQLite (local default). Each memory item is LLM-extracted from conversation turns and stored as a structured record with temporal metadata. The two-layer system maintains separate Experience (episodic, time-decayed) and Skill (procedural, durable) stores. Multi-agent isolation and per-user profiles are built in.
- Retrieval
- Four-way hybrid retrieval: dense vector similarity, full-text keyword search, knowledge-graph traversal, and recency weighting — all combined in a single query without custom glue. Memory consolidation runs `distill_all()` to merge near-duplicate Experience entries and promote recurring patterns to Skills. LLM-driven auto-merge and intelligent lifecycle management reduce retrieval noise.
- Self-host
- Self-host: moderate
- License
- Apache-2.0
- Pricing
- Open-source Apache-2.0, free to self-host. Default SQLite backend requires no external database. PostgreSQL or OceanBase backends require provisioning that database. No hosted cloud tier found. · Free / OSS
- GitHub stars
- 722
- Last release
- 2026-06-23
- Last commit
- 2026-06-25
- First catalogued
- 2026-06-28
Strengths
- Two-layer Experience + Skill distillation enables procedural memory: agents learn reusable workflows from past interactions, not just facts
- Full memory lifecycle out of the box: LLM-driven extraction, update, merge, Ebbinghaus decay — no hand-crafted pruning logic required
- Four-way hybrid retrieval (vector + full-text + graph + recency) in a single call; multi-agent isolation and user profiles built in
- Apache-2.0, actively maintained (v1.1.5, 2026-06-23); ships first-party plugins for Claude Code, OpenClaw, OpenCode, Codex, Cline
Watch out
- OceanBase ecosystem origin: the most capable backends (vector + graph together) may favor OceanBase infra; SQLite default limits some hybrid-retrieval features
- LLM-driven extraction at write time adds latency and API cost per conversation turn — profile before deploying at scale
- 722 stars and no independent benchmarks yet; self-reported benchmark harness lives in `benchmark/` but results are vendor-authored
Best for
- AI coding agents and multi-agent systems that need both factual recall and reusable procedural workflows distilled from past sessions
- Teams wanting a production-ready memory backend that spans multiple agent clients (Claude Code, Codex, OpenCode, Cline) via a shared server
Benchmark results
No sourced results yet.
Sources
- PowerMem README (vendor)
- GitHub API repo metadata (stars, Apache-2.0 confirmed via LICENSE file, v1.1.5 release) (third-party)
Last verified 2026-06-28 · updated by discover-frameworks