Mnemosyne — Cognitive Memory Engine

Give the agent a real long-term memory: four layers (raw → working → medium → long-term), scored by cognitive-psychology formulas instead of neural embeddings.

Core formula (compound-cue theory)

familiarity = 0.35·imp + 0.25·recency + 0.25·keyword + 0.10·hitFreq + 0.05·layerW

Every weight traces to a named psychology paper (Raaijmakers & Shiffrin 1981, Ebbinghaus 1885, Zeigarnik 1927). No hyperparameter sweep, no neural network.

Install

git clone https://github.com/ElonAug7/Mnemosyne-agentmemory-engine-openclaw-hermes
cd Mnemosyne-agentmemory-engine-openclaw-hermes/Mnemosyne-v6.4
bash install.sh

Installer auto-injects the memory protocol into SOUL.md / AGENTS.md and serves a Web UI at http://127.0.0.1:8765.

Protocol (mandatory)

• Before every reply: read memory/short/working/last-recall.json; if it holds flashbacks, inject them as context.
• For history/decisions/facts/preferences queries: run node tools/memory-engine/engine.js recall --query "<keywords>".
• Cite memory source in replies (e.g. "Per your decision on 2026-08-16…").

Common commands

• node tools/memory-engine/engine.js search --query "..." --mode keyword — search (~7ms)
• node tools/memory-engine/engine.js recall --query "..." — structured recall
• node tools/memory-engine/engine.js profile — user profile (tech stack, style, decisions)
• node tools/memory-engine/engine.js todos — todo list
• node tools/memory-engine/engine.js sync --quick — session maintenance

Why not embeddings

• Zero LLM API calls, zero vector DB, zero embedding models — memory stays on the machine as diffable Markdown.
• Memory-Native Evaluation (80 queries, 11 systems): nDCG 0.238, beating raw BM25 (0.185) and all embedding systems.
• Cognitive weights decide what to remember and forget — the part embeddings ignore.

References

• references/ — full architecture, benchmark methodology, Hermes adapter docs.