dsh-meow-memory: A Cross-Session Memory Plugin for DeepSeek Harness
phant0meow/dsh-meow-memory
Cross-session memory plugin for DeepSeek Harness with seven-layer structured memory and automatic injection/retrieval.
Let DeepSeek Harness remember user preferences, project details, and historical lessons across long conversations, without forgetting across sessions.
dsh plugin --profile web add github:phant0meow/dsh-meow-memory- Category
- Memory & Context
- Platform
- DSH-Plugin
- Author
- phant0meow
- Distribution
- Plugin
dsh-meow-memory Key Features
dsh-meow-memory Repository
dsh-meow-memory Repo Summary
meow-memory is a cross-session memory plugin (DSH plugin) for DeepSeek Harness (DSH), maintained by Phant0Meow, licensed under MIT, and developed in TypeScript. It solves the problem of AI assistants forgetting context in long conversations by maintaining a structured memory database (.dsh-meow/memory.db, based on node:sqlite) in each workspace, allowing the model to remember user preferences, project details, historical lessons, and other key information across sessions. Core capabilities include a seven-layer memory structure (soul, user, project, fact, lesson, topic, rules), first-turn long-term memory injection, per-message keyword matching (top-2), current project anchoring, cache-friendly design (static guide section does not break KV cache), and a nightly dream mechanism for automatic memory consolidation. The plugin also provides a complete toolset (memory_remember, memory_search, memory_project, etc.) and features collapsible injection UI to keep the message flow clean.
What core features does the meow-memory DSH plugin support?
meow-memory implements a seven-layer memory model, each layer backed by a SQLite table with time-prefixed UUIDs ensuring order. The first turn injects a fixed-format long-term memory block (full soul, full user, design principles, memory guide) and does not run keyword matching; from the second turn onward, every real user message triggers keyword matching based on entry keywords (LLM-extracted or automatic bigrams) rather than full text, with a scoring formula incorporating intersection score, IDF, coverage, Ebbinghaus decay, importance weight, and title bonus. Current project anchoring allows unanchored matches to search only global scope, avoiding false positives during casual chat. The cache-friendly design registers a static meow-memory:guide section in the system prompt with constant text, friendly to KV cache; seen memories are recorded per session and never re-injected.
How to install meow-memory?
There are two ways to install meow-memory: via npm into the profile's node_modules, or by manually copying/symlinking to ~/.dsh/profiles/web/node_modules. The recommended way is to use the DSH plugin installation command:
dsh plugin --profile web add github:phant0meow/dsh-meow-memory
After installation, add meow-memory to the dsh.profile.bundles array in the profile's package.json (recommended since v0.9.0), then restart dsh web; new sessions will load the plugin automatically.
How do the dream mechanism and reflection feature work in meow-memory?
Each window's main agent consolidates memory at night (calculated by timeZone, default 00:00–07:00), only its own memory, freezing knowledge with the last conversation timestamp. Dream deduplication is ensured via a check gate (DB atomic 60s throttle), start idempotent preemption (dream_pending), interruption self-healing, and orphan finalization. Reflection triggers after ≥7 consecutive tool steps in a single task, asking the model if there is anything worth remembering; if the last tool is memory_*, it is considered already actively remembered and no reflection is repeated.
Which DeepSeek Harness versions and platforms are compatible with meow-memory?
meow-memory has zero runtime dependencies, using Node.js built-in node:sqlite (Node ≥22.5) with no native modules, so it is compatible with any DeepSeek Harness installation running Node ≥22.5. The plugin ships with an official dsh.bundle manifest, supporting the bundle mechanism for automatic assembly. The project was last updated in 2026-08 and is actively maintained.
What is the purpose of the collapsible injection UI in meow-memory?
The client UI collapses the injected text of first-turn long-term memory and per-message keyword matches into a "▸ Injected memory" bar, while user prompts are displayed as bubbles, keeping the message flow clean. Reflection/dream turns' prompts and subsequent think/tool call/report are also collapsed into a bar, expandable to view full records with clickable details.