@hy-sde-org/dsh-memory

Agent-curated long-horizon memory for DeepSeek Harness — the ctx.memory service and its provider registry, ported from the @oh-my-pi coding-agent memory surface (the port lives in the hy-sde fork). Memory is durable, project-scoped data the agent curates itself with the retain/recall/reflect/memory_edit/learn tools (shipped by @hy-sde-org/dsh-tool-memory), and it is reloaded at the start of the next session through prompt injection. It complements DSH's session-query and compaction instead of overlapping them: those replay conversation history, this bank answers "what did we decide / prefer / learn here?" across sessions.

This is a standalone plugin build: the harness integration (the host-plane memory row) ships in cordis.patch.yml, and the agent-plane tools live in @hy-sde-org/dsh-tool-memory. Nothing in the upstream DeepSeek Harness (dsh-v0.2.0-rc.2 and later) needs to change.

Only the local backend ships. The registry keeps the seam open for Hindsight/Mnemopi-style providers later — a future provider registers one MemoryBackend and the same tools work unchanged.

Install

pnpm install --global @deepseek-ai/dsh

Direct from npm (published)

Both packages are published on the npm registry under the hy-sde-org organization (@hy-sde-org/dsh-memory and @hy-sde-org/dsh-tool-memory, version 0.2.0-rc.2). Add the service, then mount the tools via a preset:

# one command; the tool package comes in as a transitive dependency
dsh plugin --profile web add @hy-sde-org/dsh-memory @hy-sde-org/dsh-tool-memory

Then copy examples/agent-preset/ from the installed package (or this repo) to ~/.dsh/.agent-presets/<id>/ and select it in the Web UI preset picker (or dsh agent). The preset row uses @hy-sde-org/dsh-tool-memory.

From the git checkout (pre-publish / development)

git clone [email protected]:hy-sde/dsh-plugins.git
cd dsh-plugins
pnpm install
pnpm --filter @hy-sde-org/dsh-memory build

MEMORY_TGZ="$(cd dsh-memory/packages/memory && pnpm pack --pack-destination /tmp | tail -n 1)"
TOOLMEMORY_TGZ="$(cd dsh-memory/packages/tool-memory && pnpm pack --pack-destination /tmp | tail -n 1)"
dsh plugin --profile web add "$MEMORY_TGZ" "$TOOLMEMORY_TGZ"

Layout

The service is host-plane: the store is durable project data that crosses sessions, so cordis.patch.yml inserts one row into the profile; the per-session tools resolve it. Data lives under <harness home>/memories/<project>/ where <project> is an encoded absolute cwd — one memory root per project, shared by every session and tool on it.

Each project root holds three artifacts:

• bank.jsonl.zstd — editable working entries written by retain (id, content, context, source, importance, timestamps, active flag). Backs memory_edit. On-disk container is the same zstd frame format the harness session logs use: each save batch is one checksummed Zstandard frame (concatenated, append-only, self-healing). The pre-rename plaintext bank.jsonl is still read and is migrated on the first write; set compression: 'none' in LocalMemoryConfig for the original line-append format. The codec ships inside this package (/frame-codec subpath, node:zlib-only).
• learned.md — newest-first, deduped, capped (100) lesson bullets written by learn; the same format and normalization omp keeps. Survives consolidation; learn writes are injection-neutralized and secret-redacted.
• memory_summary.md — optional consolidated summary (hand- or tool-maintained) that recall, reflect, and prompt injection surface.

Service API

const memory = ctx.memory                       // MemoryService
await memory.save({ cwd }, { content, context, source, importance })
await memory.learn({ cwd }, { content, context })
await memory.search({ cwd }, 'query', { limit: 10 })
await memory.edit({ cwd }, 'update' | 'forget' | 'invalidate', { id, content, importance, replacementId })
await memory.summaries({ cwd })                 // { summary?, learned?, block }
await memory.status({ cwd })
await memory.clear({ cwd })

Backends register into the service: memory.register(backend) returns a disposer, and memory.resolve() returns the configured (default: first) backend. Mutations emit memory/change ({ cwd }) so in-process consumers can invalidate caches.

A MemoryBackend is a dozen methods over @hy-sde-org/dsh-memory/types; the shipped one, LocalMemoryBackend, is pure-node (node:fs) with an in-process per-file write chain so concurrent saves from sibling sessions can never drop each other's writes.

memory:// internal URLs

When the host composition mounts @hy-sde-org/dsh-internal-urls (the shared registry behind the read/grep tools), this plugin registers a memory:// scheme in ctx.internalUrls — once per process, via ctx.inject, and it is a graceful no-op in compositions without the registry. Every resolved resource is immutable: true: agents never rewrite durable memory through a file-shaped URL; memory_edit is the mutation surface.

Two URL forms:

• memory://root — the project's consolidated memory overview (summary + learned lessons + working bank, the same block prompt injection uses). An empty project reads a "Project memory is empty" pointer instead.
• memory://<id> — one stored entry in full, with a metadata header (id, source, importance, timestamp, readonly). Ids are the same ones recall/reflect surface: bank rows by their m_* id, lessons as lesson_<hash> (read-only), and the consolidated summary as summary_0 (read-only). Reads are scoped to the calling session's project (cwd).

Corrective errors, following the upstream oh-my-pi (coding-agent/src/internal-urls/memory-protocol.ts) HINDSIGHT_UNADDRESSABLE pattern:

• not addressable — the selected backend has no readEntry (a non-addressable store): "The … memory backend is not addressable via memory://<id>", pointing back at recall/reflect.
• not found — the id does not exist in this project (or was retired): "Memory … does not exist in this project", pointing at recall/memory_edit.
• a missing namespace, a missing cwd, an unregistered backend, and paths under root/an id all get their own corrective messages.

Completions: complete() returns root plus every addressable entry id (listEntries, newest first, capped at 50) with a one-line content preview; a backend that only implements readEntry falls back to a <id> placeholder, and one with neither completes only root. The memory:// scheme never takes a path (backend-shaped, not file-shaped).

Config (the memory row)

Per-entry caps: bank content 4000 chars, bank context 800, lesson content 2000, lesson context 400, lessons capped at 100 newest-first. All stored text passes injection-neutralization (control chars, </backticks, ~~~ fences) then secret-redaction, on write and on read.

Tests

pnpm -r --filter @hy-sde-org/dsh-memory test