dsh-engram-recap

Forces periodic Engram memory recall and save reminders for DeepSeek Harness agents, so knowledge survives across sessions without depending on the model remembering to do it unprompted.

Problem

Engram's own MCP tool descriptions say things like "call this PROACTIVELY — don't wait to be asked", but that is advisory prose in a tool description: nothing in the harness enforces it. In practice, an agent that is not explicitly told to save memory tends not to, and when a session ends without a save, everything learned in it is lost.

What it does

Mounts a dynamic systemPrompt context (re-evaluated on every model request, unlike a static persona section read once at session start) that:

1. Turn 0 — reminds the agent to call mcp__engram__mem_context (and mem_search for anything specific) before its first substantive action.
2. Every remindEvery turns (default 6) — reminds the agent to call mcp__engram__mem_save if the recent exchanges produced a decision, bugfix, discovery, or lesson learned. If nothing memory-worthy happened, the reminder tells it to do nothing (no forced noise-saving).

The reminder is injected as a <system-reminder> block, the same framing convention dsh-agent-instructions uses for workspace instructions.

Why a prompt nudge and not a direct tool call

The Host plugin execution environment has no AbortController or crypto available, so this plugin cannot safely drive ctx.tools.execute(...) against Engram's MCP tools (they require an AbortSignal). Steering the model via prompt context — the same mechanism dsh-plan-mode uses for its policy section — is the correct approach given that constraint, and it keeps the actual judgment (what is "memory-worthy") with the model instead of a rigid heuristic.

Configuration

- id: engram-recap
  name: 'dsh-engram-recap'
  config:
    remindEvery: 6   # optional, positive integer, default 6

Requirements

Requires an Engram MCP server (mcp__engram__* tools) mounted on the agent; this plugin only nudges toward those tool names, it does not provide them. With no Engram tools available, the reminders are harmless no-ops from the model's perspective (it simply cannot act on them).

State and lifetime

Per-session turn counts live in an in-memory WeakMap keyed by Session, process-local and reset on restart — same lifetime tradeoff as any other stateless per-request Cordis contribution.

Install

dsh plugin --profile web add dsh-engram-recap

Or via the homologated stack in dsh-manage — see its plugins/manifest.json.