DeepSeek Harness plugin

agent-memory-system

DeepSeek Harness community Bundle for local, evidence-backed Agent Memory

Jump to install

Source facts

Repository
rrrrrredy/agent-memory-system
Latest update
Aug 18, 2026
Category
Workflow & Automation
GitHub stars
0
Format
plugin
Package path
integrations/deepseek-harness
Catalog evidence
Upstream dsh.bundle evidence
Evidence path
integrations/deepseek-harness/package.json#dsh.bundle
Checked against
0.1.0-rc.8
Upstream check date
2026-08-20

This evidence comes from the upstream catalog. This site has not installed, run, or security-reviewed the plugin.

Install

Start with a prompt that asks an agent to review the GitHub repository and source. Switch to the command if you want to install it yourself.

Copy this prompt into DSH, Codex, or another agent and ask it to review the GitHub repository and source first.

Do not install or run any commands yet. Read this plugin's GitHub repository, README, and relevant source code. Then answer the questions below clearly and directly so I can decide whether it fits my needs:

1. What is this plugin, and what problem does it solve?
2. Who is it for, and what are its typical use cases?
3. How is it used after installation? Include one minimal example.
4. What known limitations or privacy, security, compatibility, or maintenance risks does it have?
5. Give a clear recommendation: recommend, conditionally recommend, or do not recommend, with reasons.

Distinguish statements documented by the repository, inferences from source code, and unknowns. If evidence is insufficient, say so explicitly. Do not guess or simply repeat the README.

GitHub: https://github.com/rrrrrredy/agent-memory-system/tree/HEAD/integrations/deepseek-harness
Plugin: agent-memory-system
Author: rrrrrredy

Check the source files

Read the README and other files from this plugin directory before installing.

File explorer3 files
README.mdSource · read only

@rrrrrredy/dsh-agent-memory

Community DeepSeek Harness Bundle for Agent Memory System. It can preserve Harness sessions as a crash-safe local append-only event spool and can inject only verified, active, promoted memory before an agent step.

This is a community plugin, not an official DeepSeek plugin.

Compatibility

  • DeepSeek Harness / @deepseek-ai/dsh: 0.1.0-rc.6
  • Agent Memory CLI: >=0.3.1 <1 for deepseek-harness import and injection support
  • Node.js: 22.19.x or 24+

Capture and retrieval are independent, explicit opt-ins. The default Bundle row is inert until at least captureDir or both retrieval roots are configured.

Install

Install the Agent Memory CLI when retrieval or import is required:

go install github.com/rrrrrredy/agent-memory-system/cmd/agentmem@v0.3.1

Install the public Bundle into the Harness profile you use:

dsh plugin --profile headless add @rrrrrredy/dsh-agent-memory@0.1.0
dsh --profile headless --dump-config

For the browser surface, replace headless with web. For a local release candidate, replace the npm specifier with the absolute path to the packed .tgz.

Configure

Update the agent-memory row in that profile's cordis.patch.yml. All storage paths must be absolute.

- id: agent-memory
  config:
    binary: agentmem
    captureDir: /absolute/local/path/outside-any-git-worktree/dsh-capture
    evidenceRoot: /absolute/local/path/agent-memory-evidence
    portableRepo: /absolute/path/to/portable-memory-repository
    scopeRepository: account/example
    scopeProject: example
    scopeTask: agent-work
    limit: 5
    tokenBudget: 600
    byteBudget: 3072
    timeoutMs: 3000
    stdoutMaxBytes: 131072
    stderrMaxBytes: 131072
    segmentMaxBytes: 33554432
  • Set only captureDir for capture without retrieval.
  • Set evidenceRoot and portableRepo together for retrieval without capture.
  • Set all three paths to compose both capabilities.
  • Trusted repository, project, and task scopes come only from this local configuration. Model text cannot select them.

On Windows, use fully qualified absolute paths and escape backslashes when the YAML parser requires it. The capture directory is rejected if it is inside a Git worktree, including through a linked path.

Use

Run Harness normally:

dsh --profile headless "Continue this repository task using only relevant verified memory."

At the first accepted step of each turn, the Bundle sends only user-originated text to:

agentmem inject deepseek-harness --root <evidenceRoot> --repo <portableRepo> ...

A successful bounded response becomes a formal user/message whose source is plugin: agent-memory and form is recall, so it remains visible and replayable in the Harness event log. Missing CLI, timeout, output overflow, nonzero exit, damaged JSON, failed verification, or an empty result all continue without memory. The Bundle never falls back to stale, partial, unverified, or cached context.

Import the local capture spool into the Agent Memory evidence ledger explicitly:

agentmem import deepseek-harness-events \
  --root <local-evidence-directory> \
  --path <captureDir>

Live session/event records are retained exactly as delivered by Harness. On startup, a raw-artifact-capable persistence backend is also read through listSnapshots and readRaw; its full session artifact is preserved verbatim. Unknown events, unsupported raw artifacts, malformed artifact rows, recovery problems, and read failures remain explicit gaps rather than disappearing.

Privacy and security boundary

  • Capture files are local-only raw evidence and may contain prompts, model output, locally exposed reasoning, tool calls, tool results, paths, and secrets produced during the session.
  • Raw Harness events, reasoning, tool output, and session artifacts must never be committed to Git, packed into npm, or attached to a GitHub Release.
  • The Bundle adds no telemetry, hosted service, remote storage, or model tool. Harness still sends ordinary conversation content to the model provider configured for the profile.
  • Retrieval returns only memory that Agent Memory validates as active and promoted from a complete portable repository. Capture does not automatically promote anything.
  • raw_exposed means the text existed in the local Harness artifact. It does not claim access to provider-hidden reasoning.
  • Local files are not an OS security boundary. A process with the same user privileges can alter local configuration or evidence.

Uninstall

dsh plugin --profile headless remove @rrrrrredy/dsh-agent-memory
dsh --profile headless --dump-config

After removal, the Bundle row and package reference are absent, no event or pre-step listener remains, and later sessions do not add capture records or memory recall messages. Existing local capture and evidence files are deliberately retained; remove them separately only under your own retention policy.

Reproduce the tests

From this directory:

pnpm install --frozen-lockfile
pnpm verify
AGENTMEM_BINARY=/absolute/path/to/agentmem pnpm test:integration
pnpm pack --pack-destination .pack

pnpm verify covers config boundaries, canonical-source hashing, Git-worktree rejection, durable raw event capture, raw-artifact backfill and restart deduplication, explicit gaps, argv-only subprocess execution, timeouts, output bounds, malformed responses, trusted prompt selection, formal recall messages, real Cordis/SessionStore/JSONL Persistence composition, and clean disposal.

pnpm test:integration writes a real Bundle spool into an explicit scratch root, imports it with the compiled Agent Memory CLI, and requires a clean agentmem doctor result. Set AGENT_MEMORY_DSH_TEST_TMP to keep all scratch work on a chosen volume.

The [0.1.0 public release evidence](release-evidence/0.1.0.md) records the separate isolated-profile public install/remove test and real DeepSeek model smoke. Raw session and memory evidence are excluded from the package and release assets.