dsh-memory

August 15, 2026 · View on GitHub

Self-learning long-term memory for DeepSeek Harness — a historian that compresses your context, a dreamer that curates what matters, and a sidekick that recalls it exactly when you need it.

English | 中文

Long-term memory for DeepSeek Harness, modeled on magic-context (capture → consolidate → recall). A single host-plane plugin; the store is a local SQLite database with FTS5 full-text search and local ONNX embeddings.

What it does

Capture — every N turns, a cheap model distills the conversation into durable memories (throttled, never blocks the main flow).
Historian — when the context window fills up, old messages are folded into tiered summaries; the original session log stays append-only and can always be recovered.
Recall — before each turn, project-relevant memories are retrieved and injected into the surface as a <system-reminder> block.
Dreamer — scheduled background maintenance: classify, merge, archive, and promote memories that matter.
Sidekick — on-demand retrieval augmentation (memory_aug): search memory to enrich the current answer.
13 memory tools — the model can write, search, list, update, archive, merge, get, delete, dream, consolidate, expand, and augment memories, plus session notes.

Prerequisites

deepseek-harness installed from source — this plugin targets the DeepSeek Harness plugin API: https://github.com/deepseek-ai/deepseek-harness.
Node.js ≥ 22 — the store uses Node's built-in node:sqlite.

Install

Clone (or symlink) this repository into the plugin directory of your harness profile:
```
<dsh-home>/profiles/<profile>/plugins/dsh-memory
```
Run npm install inside the plugin directory.

Reference it from the profile composition:

- id: dsh-memory
  name: dsh-memory
  config: {}

On first load the database is created and migrated automatically — no manual setup.

Built-in compaction

dsh-memory's historian takes over context folding, so the host's built-in compaction should be turned off in the agent preset you use — the same idea as magic-context's setup wizard, which disables the host compressor at the config layer. scripts/setup.mjs toggles the compaction-basic / command-compact entries of any preset in <dsh-home>/.agent-presets/ with point edits only (comments and formatting are preserved):

node scripts/setup.mjs enable  --preset <name>   # host compaction off → historian owns folding
node scripts/setup.mjs disable --preset <name>   # undo only what this script added
node scripts/setup.mjs status                    # per-preset compaction state

--home defaults to ~/.dsh; --preset all applies to every preset that has the compaction entries. Each enable backs up the preset yml (agent.cordis.yml.dsh-memory.bak) and records what it changed in .dsh-memory-setup.json, so disable never touches entries the user disabled by hand.

Tools

All tools are model tools: the model calls them through the harness.

Tool	One-liner
`memory_write`	Write a memory; idempotent dedup
`memory_search`	Hybrid retrieval across memories, compartments, and session notes
`memory_list`	Filter by category / scope / status, with pagination
`memory_update`	Edit a memory by id; re-embeds and re-classifies
`memory_archive`	Soft delete (auditable) — prefer over delete
`memory_merge`	Merge duplicates; keep the most important row
`memory_get`	Read memories by id
`memory_delete`	Hard delete (last resort)
`memory_dream`	One manual consolidation run: merge + classify + curate
`memory_consolidate`	Merge cosine duplicates; clean up expired rows
`memory_expand`	Recover the original text of a compressed partition
`memory_aug`	Sidekick augmentation synthesis
`ctx_note`	Session-scoped notes (write / read / dismiss / update)

Configuration

Every field is optional; out-of-range numbers are clamped, invalid values throw at load. The most common knobs:

config:
  embed:
    mode: local          # local | openai-compatible | none
  capture:
    everyTurns: 5        # distill the conversation every N turns
  inject:
    topK: 4              # memories recalled per turn
  historian:
    triggerRatio: 0.70   # compress when ~70% of the window is used
  dream:
    enabled: false       # scheduled background maintenance (default off)

See lib/config.js for the full field list, defaults, and clamping rules.

Acknowledgments

This is an independent reimplementation for deepseek-harness, inspired by and partially ported from magic-context by ualtinok (MIT License). Not affiliated with cortexkit or DeepSeek.

Ported / derived files (per code comments): lib/redaction.js, lib/cron.js, lib/decay.js, lib/dreamer.js, lib/budgets.js; other modules are modeled on magic-context's design. See NOTICE for the upstream license text.

License

MIT — see LICENSE. Upstream notices in NOTICE.