dsh-memoria

Long-term memory for DeepSeek Harness, backed by the Memoria memory framework.

The plugin owns one long-lived Python subprocess (python -m memoria.plugin_server) so store and recall share a single in-memory Memoria instance. It registers four tools and — by default — auto-injects recalled context at the start of every turn, so the agent stays grounded in what it already remembers without an explicit recall prompt.

Requirements

• A Python interpreter that can import memoria (Memoria installed into that interpreter), or a cwd that contains the memoria package.
• The python and cwd defaults are portable (python/python3 on PATH, process.cwd()); override them per machine (see Configuration).

Install

dsh plugin --profile <profile> add dsh-memoria

Or, for a local checkout:

dsh plugin --profile <profile> add link:/path/to/memoria-dsh-plugin

The package declares a dsh.bundle patch, so dsh plugin add reconciles it into the profile's dsh.profile.bundles automatically.

Configuration

Set these in the profile's cordis.patch.yml:

- id: memoria
  config:
    python: /path/to/python            # interpreter that can import memoria
    cwd: /path/containing/memoria      # subprocess working dir
    autoInject: true                   # inject recall at agent/pre-step
    recallLimit: 10                    # max memories injected per turn (1-50)

Key	Default	Description
python	python (win) / python3 (other)	Interpreter for the Memoria subprocess.
cwd	process.cwd()	Working directory of the subprocess; must let python -m memoria.plugin_server resolve.
autoInject	true	Whether to auto-inject recall before the first step of each turn.
recallLimit	10	Maximum reference memories injected per turn.

Tools

• memoria_store — store a durable, context-free fact (content, optional types, importance, tags).
• memoria_recall — recall structured context against a query.
• memoria_forget — remove one memory by memory_id.
• memoria_status — report per-layer memory counts and lifecycle distribution.

How it works

apply(ctx, config) spawns python -m memoria.plugin_server over stdio (newline-delimited JSON-RPC) and registers the four tools on ctx.tools. A prepend listener on agent/pre-step calls recall for the turn's text and, when it returns memories, prepends them as a user message tagged source.form = "recall" — so the model sees its memory before answering, and the injected message never reaches the model as tool chatter.