Phant0Meow/dsh-memory-meow

Project-scoped cross-session memory: six-layer SQLite store (soul/user/project/fact/lesson/topic) on node:sqlite, cache-friendly first-message injection, memory_remember/search/find_similar/read/update tools, and per-window nightly consolidation ("dream").

Cross-session memory for DeepSeek Harness: every workspace keeps a structured SQLite memory database (.dsh-meow/memory.db). Seven memory layers (soul / user / project / fact / lesson / topic / rules), one table per layer with time-prefixed UUIDs. The static tool manual lives in the system prompt as a fixed section — constant text, so LLM provider KV/context caches stay untouched. Dynamic content (soul/user in full, design principles, memory guide) is injected as a prefix of the first user message; the first turn injects long-term memory only (no keyword hits). From the second user message on, every message gets a keyword hit (top-2) scored by intersection × idf × coverage × Ebbinghaus decay × importance × title bonus. The model deep-dives with memory_search / memory_project on demand, and each window's agent consolidates its memories while idle ('dream'). Cache-friendly by design: already-seen memories are recorded per session so injection never repeats; a session-compaction signal releases them so compressed-away memories can be hit again.

Memory ★ 56 updated 2026-08-28
View on GitHub ↗

Install

cd ~/.dsh/profiles/web && npm install meow-memory && dsh web

npm package meow-memory 0.20.0 (registry-verified 2026-08-28): install into the profile's node_modules (cd $DSH_HOME/profiles/web && npm install meow-memory), add meow-memory to the profile package.json's dsh.profile.bundles array (the package ships a dsh.bundle.patch; recommended since v0.9.0), then restart dsh web. Manual fallback: symlink the package into ~/.dsh/profiles/web/node_modules/meow-memory and add it to bundles the same way.

Compatibility

DeepSeek Harness web profile. SQLite via node:sqlite (Node built-in). Workspace-scoped (.dsh-meow/memory.db).

Details

Recent updates

seven-layer SQLite memory; KV-cache-friendly system-prompt guide; first-turn long-term injection; per-message keyword hits; idle 'dream' consolidation.

FAQ

Does it break LLM prefix caching?
No — the static tool manual is a fixed system-prompt section (constant text) and memory enters history like normal messages, so the prefix stays identical and KV cache hits are preserved.
What are the seven layers?
soul (the AI itself), user (basics & preferences), project (per-project info with subcategories), fact (atomic facts), lesson (mistakes & corrections), topic (ongoing discussion arcs), rules (design principles).
How do I install it?
cd ~/.dsh/profiles/web && npm install meow-memory, add meow-memory to the profile package.json's dsh.profile.bundles, then restart dsh web.

Alternatives

Scorp1o117/dsh-tdai-memory · Aik358/dsh-auto-memory · Towzai/dsh-memory

More plugins in Memory

Browse more in Memory

Guides for Memory plugins