Aloneswork/deepseek-harness-evolving-memory
DeepSeek Harness Local semantic evolving long-term memory plug-in｜Local semantic evolving memory for DSH
deepseek-harness-evolving-memory gives DeepSeek Harness automatic, local long-term memory across conversations, workspaces, and projects: it saves reusable successful results and recalls relevant memories before DSH answers, using real local BGE embeddings so a reworded question can still match. Memory is captured automatically after successful turns (combining embeddings, lexical relevance, confidence and time decay) and recalled during DSH's first agent/pre-step without the model needing to call a search tool. A stable memory key updates identical facts instead of duplicating them, uncertain changes go to a conflict queue, and a replacement must exceed the old confidence by at least 0.15 to supersede it. It supports confidence, expiry, version history, project isolation, archive, secure delete, and Markdown/JSON export, and migrates v0.1 SQLite data to v0.2 in a transaction. Explicit MCP tools remain available (memory_save, memory_search, memory_get, memory_revise, memory_archive, memory_delete, memory_conflicts, memory_conflict_resolve, memory_export).
Install
curl -LO https://github.com/Aloneswork/deepseek-harness-evolving-memory/releases/download/v0.2.1/deepseek-harness-evolving-memory-0.2.1.tgz && npm install --global ./deepseek-harness-evolving-memory-0.2.1.tgz && dsh plugin --profile web add ./deepseek-harness-evolving-memory-0.2.1.tgz && dsh web --patch "$(npm root --global)/deepseek-harness-evolving-memory/evolving-memory.cordis.yml"Requirements: Node.js >= 22.19.0, DeepSeek Harness installed and configured, and about 91 MB of disk space for the local BGE-small-zh-v1.5 model downloaded on first use. The README downloads the v0.2.1 release tarball, installs it globally, adds it to the web profile, and starts DSH with the plugin's cordis patch; the last command enables it for that DSH Web process -- to enable permanently, merge the two insert entries from evolving-memory.cordis.yml into your profile or global cordis.patch.yml (do not overwrite existing entries). No registry package is claimed; the README also documents a from-source path (git clone, npm install, npm test, npm pack --dry-run).
Compatibility
Node.js >= 22.19.0; DeepSeek Harness installed and configured. Uses the official @deepseek-ai/dsh-mcp-client for MCP child-process reconnect and tool rediscovery. About 91 MB of disk space is needed for the local BGE-small-zh-v1.5 embedding model, downloaded on first use. Memory text and embeddings stay in a local SQLite database; the README states no memory text is sent to an embedding API. First model download contacts the model distributor but uploads no memory content.
Details
- Repo: Aloneswork/deepseek-harness-evolving-memory
- Category: Agent Capabilities
- Stars: 1
- Version: GitHub release tarball v0.2.1 (no npm registry package claimed); MIT
- Last push: 2026-08-14
- First seen: 2026-08-14
Recent updates
The README documents the storage locations (~/.dsh-evolving-memory/memories.sqlite for the database and ~/.cache/dsh-evolving-memory/models for the model cache), the default recall scope (global plus the current project; cross-project recall only with includeAllProjects=true), project identity resolution order (explicit tool argument, EVOLVING_MEMORY_PROJECT, Git origin, then the current directory), the optional EVOLVING_MEMORY_FILE / EVOLVING_MEMORY_PROJECT / EVOLVING_MEMORY_MODEL_CACHE variables, the native plugin config keys (autoRecall, recallLimit, autoCapture, captureMinChars, captureMaxChars, captureExpiresDays), the confidence rules, and the privacy boundary. A dsh-evolving-memory --doctor command is documented for verification.
FAQ
- How do I install DeepSeek Harness Evolving Memory?
- Per the README: download the v0.2.1 release tarball with curl, npm install --global it, dsh plugin --profile web add it, then start dsh web --patch "$(npm root --global)/deepseek-harness-evolving-memory/evolving-memory.cordis.yml". Node.js >= 22.19.0 and a configured DeepSeek Harness are required.
- Where is my memory stored?
- The README states memories and vectors live in a local SQLite database at ~/.dsh-evolving-memory/memories.sqlite and the embedding model cache at ~/.cache/dsh-evolving-memory/models; no memory text is sent to an embedding API.
- Does it recall memories automatically?
- Yes -- the README says relevant memories are recalled during DSH's first agent/pre-step so the model does not need to remember to call a search tool, and reusable request+result pairs are captured automatically after successful turns (default confidence 0.55, expiring after 180 days).
Alternatives
Aik358/dsh-auto-memory · csyangwen/dsh-memory-evolve · djasdh/interest-memory