dsh-memory-tencentdb

English | 中文

DeepSeek Harness (dsh) ↔ TencentDB Agent Memory (TAM) long-term memory bridge plugin.

A Cordis plugin that gives DeepSeek Harness Web UI sessions a read-write loop with the TAM four-layer memory system (L0 conversation → L1 extraction → L2 scenes → L3 persona). Mechanism is aligned with the Hermes memory_tencentdb provider: from TAM's perspective a dsh session is just another agent session — same v3 data-plane API, same identity tenancy, same record shape.

About

DeepSeek Harness is an agent harness with a Web UI where you talk to an agent. TAM (TencentDB Agent Memory) is a team-level long-term memory service that distills conversations into reusable memories. This plugin connects the two:

• Write: every finished turn is captured into TAM L0 (user + assistant), and the TAM pipeline automatically distills L1/L2/L3 — no extra work on the dsh side.
• Read: before every model call, the plugin recalls relevant L1 memories, the L3 persona and L2 scene navigation from TAM and injects them as <memory-context>, so the agent "just remembers" past conversations and team doctrine.

Because it reuses the same identity as your existing agents (e.g. Hermes), memories accumulated from dsh conversations are visible to the rest of the team and vice versa — one unified memory view across tools.

Features

• L0 conversation capture (write) — per-turn user+assistant pairs via POST /v3/conversation/add; asynchronous, never blocks the agent loop.
• Recall injection (read) — on every prompt assembly, parallel atomic/search (L1) + core/read (L3) + scenario/ls (L2), rendered into a <memory-context> block identical to Hermes' build_memory_context_block.
• Connection lifecycle — probe on startup, reuse if running, optional auto-start if not (Hermes supervisor.py semantics), watchdog keep-alive, circuit breaker.
• Removable without side effects — pure HTTP client; removing the plugin never touches the TAM gateway process or accumulated data.
• Silent degradation — gateway down: writes are skipped, recall is skipped, the conversation is never blocked.

How it works

dsh Web UI session
 ├─ write: turn/end ──► POST /v3/conversation/add  (user + assistant, ISO-8601)
 ├─ read:  system-prompt/assemble ──► atomic/search + core/read + scenario/ls
 │         └─► <memory-context> injected into the assembled prompt
 └─ lifecycle: health probe → reuse → (optional) spawn → watchdog → dispose-only-own-child

All data-plane calls go to the same TAM gateway (default http://127.0.0.1:8420) with Bearer auth and x-tdai-service-id: default, exactly like Hermes.

Installation

# 1. Install into a profile (web / headless)
dsh plugin --profile web add /path/to/dsh-memory-tencentdb

# 2. Mount in the profile's cordis.patch.yml
- insert:
    - id: dsh-memory-tencentdb
      name: '@dsh-local/dsh-memory-tencentdb'
      config:
        enabled: true

# 3. Restart dsh
systemctl restart dsh-web   # or however you run dsh

Configuration

All connection parameters come from environment variables (env first, config fallback, defaults match Hermes):

Env var	Purpose	Default
MEMORY_TENCENTDB_GATEWAY_HOST / _PORT	TAM gateway address	127.0.0.1:8420
MEMORY_TENCENTDB_GATEWAY_API_KEY / TDAI_GATEWAY_API_KEY	gateway Bearer key (takes precedence over keyFile)	—
MEMORY_TENCENTDB_TEAM_ID / AGENT_ID / USER_ID	TAM tenancy identity (share the same values as Hermes for a unified memory view)	default
MEMORY_TENCENTDB_GATEWAY_CMD	gateway start command (used only for optional auto-start)	auto-discovered

Fallback key file: ~/.memory-tencentdb/biz-user-key.txt (600 perms, first non-empty line).

The plugin config supports enabled / granularity / recallEnabled / writeOnAbort etc. — see the Config schema in lib/index.js.

Removal

# delete the insert block from cordis.patch.yml
dsh plugin --profile web remove @dsh-local/dsh-memory-tencentdb
# restart dsh

The TAM gateway service and all accumulated memories are unaffected.

Alignment with Hermes

Aspect	Value
Write endpoint	POST /v3/conversation/add (same body, two messages per turn)
Read endpoints	atomic/search (L1), core/read (L3), scenario/ls (L2)
Identity	same team_id / agent_id / user_id resolution (env → config → default)
Injected text	<memory-context> block, byte-aligned with Hermes
Reliability	circuit breaker 5/60s, single-attempt writes, watchdog 10s
Session end	handled by the TAM pipeline automatically

Compatibility

• DeepSeek Harness 0.1.0-rc.* (event/service signatures follow the installed .d.ts)
• TAM Gateway v3 data-plane API
• Zero external runtime dependencies (Node built-in fetch)

License

MIT