djasdh/interest-memory
Low-footprint memory backend for AI agents — single binary, ~50MB RAM, verify-augmented accuracy
A long-term memory backend for AI agents: one ~50MB process replaces a Postgres + Redis + vector-DB stack. At session end it extracts interest points from the transcript, verifies and cleans them, and writes them into a local knowledge base as wiki pages via an LLM-driven agent loop; at session start it recalls and injects relevant context. Semantically similar interest points are merged automatically so memory converges with use instead of bloating. Multiple agents can share one service with isolated, fully-shared, or selective sharing, and every structural change is written to a replayable change_log.
Install
dsh plugin --profile web add @djasdh/interest-memory-dsh-bridgeTwo parts: (1) the memory daemon — a single ~18MB binary + SQLite (one-command curl install per the README; LLM/embeddings can point at local Ollama/vLLM for fully offline use); (2) the DSH bridge — dsh plugin --profile web add @djasdh/interest-memory-dsh-bridge (npm @djasdh/interest-memory-dsh-bridge 0.1.1, verified npm 2026-08-31; source bridge/dsh/). The bridge provides session-start recall injection, session-end transcript push, and memory_search / memory_logs consumer tools.
Compatibility
DSH web profile; standalone daemon (Linux/macOS, runs on Raspberry Pi, ~17MB idle / <75MB peak); works with Hermes / OpenCode / Claude Code / Codex bridges too; SQLite + optional sqlite-vec.
Details
- Repo: djasdh/interest-memory
- Category: Agent Capabilities
- Stars: 2
- Version: npm @djasdh/interest-memory-dsh-bridge 0.1.1
- Last push: 2026-08-21
- First seen: 2026-08-02
Recent updates
Session-end interest extraction → verify → wiki-page knowledge base; session-start recall injection; memory_search/memory_logs tools; auto-merge of similar points; multi-agent sharing; replayable change_log.
FAQ
- How heavy is it?
- One ~18MB binary plus one SQLite file is the entire footprint — ~17MB idle and <75MB peak measured, so it runs on a Raspberry Pi with no external databases or cloud dependency.
- Can it run fully offline?
- Yes — LLM and embedding endpoints can point at local Ollama/vLLM, keeping every byte of memory on your machine.
- Do multiple agents share the same memory?
- Yes — one daemon serves many agents (DSH, Hermes, OpenCode, Claude Code, Codex) with isolated, fully-shared, or selective sharing per agent.
Alternatives
Phant0Meow/dsh-memory-meow · vpromise/dsh-memory · Scorp1o117/dsh-soul-md