v0.2.0 — first public release

First public release. 🌊

Memory River is a local-first, long-term memory engine for AI agents: a session recap that fades, distilled facts that persist, and a verbatim transcript the agent can always go back and re-read — with a health model where memories decay, heal on access, and get superseded.

Highlights

• Three-timescale memory — short-term capsule / medium-term notes / long-term transcript, every distilled claim traceable back to its source turns (sourceEntryIds → rehydrate)
• Two-pass retrieval — cheap automatic recall + precise on-demand rehydrate by entry-id / time window / keyword
• Memory that metabolizes — forgetting-curve health, supersede chains, conflict deprecation, nightly consolidation
• One engine, many doors — framework-neutral core (createMemoryRiver), OpenClaw adapter, MCP stdio server, Claude Code push-layer plugin, loopback HTTP daemon (mr-serve), read-only web dashboard
• 3-question onboarding — mr init wizard + mr doctor seven-point environment check; no-LLM-key degraded mode keeps archiving and recall working
• Honest durability semantics — WAL with idempotent at-least-once roll-forward (SIGKILL-tested), single source of truth on the durable tier, bounded transcript retention; see the README's Durability & operational honesty section
• Benchmarked in the open — full LoCoMo (1,986 questions, end-to-end through a real agent): 60.6% strict / 79.9% mem0-style rubric on cat 1–4, full protocol documented in the README

Requirements

Node 22+, local Ollama for embeddings. CI covers Node 22 and 24.

Notes

• APIs may still move before 1.0.
• Not yet published to npm — install from source for now (see Quick Start in the README).

License: Apache-2.0