dsh-plugin-hindsight-advanced

A plugin for DeepSeek Harness (dsh) that integrates the Hindsight memory system: your agent stores memories in your own Hindsight server and brings the relevant ones back into the conversation.

Installation

With the dsh command:

dsh plugin --profile web add /path/to/hindsight-advanced

When running dsh from source:

pnpm dsh plugin --profile web add /path/to/hindsight-advanced

Then restart the server.

Configuration

The configuration is one hindsight row in your own DSH config — never an edit to the plugin's files. You place it in one of two spots (pick one, never both):

Every session. In your profile's patch layer (~/.dsh/profiles/web/cordis.patch.yml, append to your existing file):

- id: hindsight
  disabled: false
  config:
    bank: my-bank
    baseUrl: http://127.0.0.1:9177
    apiKeyRef: HINDSIGHT_API_KEY

One preset only. In that preset's own file (~/.dsh/.agent-presets/<preset>/agent.cordis.yml, append):

- id: hindsight
  name: dsh-plugin-hindsight-advanced
  config:
    bank: my-bank
    baseUrl: http://127.0.0.1:9177
    apiKeyRef: HINDSIGHT_API_KEY

Presets that point at the same bank don't need separate banks — the preset scope already keeps each preset's memories apart.

Two things to know about the row:

• the config block replaces the plugin's defaults wholesale — list every key you want; only bank is strictly required, the rest fall back to the defaults in the table below (but baseUrl and apiKeyRef are worth repeating, since their defaults point at port 8888 and carry no key);
• to switch the plugin off again, set disabled: true on the row (or remove the preset row) and restart.

The configuration keys

key	default	what it does
bank	—	Required. The name of the Hindsight bank to use (the server creates it on first use).
baseUrl	http://127.0.0.1:8888	Your Hindsight server's address.
apiKeyRef	—	The way to give the plugin its key — a name for the secret, not the secret itself. The value is looked up on every call (environment → ~/.dsh/.credentials.yaml → .env files), so rotating the key needs no restart and the key never sits in a config file. See below.
apiKey	—	A plain key in the config, instead of a reference. Prefer apiKeyRef.
autoContext	true	Set false to turn off the automatic per-turn recall (the hindsight tool stays).
prefetch	true	true (default): the recall starts as the previous turn ends, so from the second turn on it targets the previous message. false: every turn queries the message you just sent and waits on the server for it — the memory is always relevant to what you just said, at the cost of the server's latency on every first model call.
recallContextTurns	5	How many of your recent turns ride along in the recall's query, so the bank can match memories against what the conversation was about, not just the last message. The query is the anchor message under a Prior context: block of the recent prior turns (one line per user/assistant message), capped at 1000 characters with the oldest lines dropped first. 1 (what the reference integrations ship) is the plain single-message query.
retainScope	preset	Where a stored memory lands when the agent doesn't say: global (everything in the bank), preset (this agent preset), or session (this session only).
maxRecallTokens	1024	How much memory to bring back per recall.
autoContextTimeoutMs	2500	How long a turn may wait on the memory server before moving on without it. Because the lookup runs ahead (below), most turns never pay this; a slow or stopped server never blocks the agent.
retainAsync	false	false (the default): storing waits until the bank has processed the memory, so the next turn already sees it. true: store acknowledges fast and the bank processes it in the background.
bankConfig	—	Optional instructions for the server's own memory extraction — for example retain_mission: "Focus on decisions and durable project facts." tells the Hindsight server what to pull out of what the agent stores.

How the automatic recall works

At the start of each turn the agent gets a small note: the memories that match your latest message, plus any standing rules you stored. The lookup starts in the background as soon as the previous turn ends — while you are reading or typing — so by the time the agent starts the next turn the memory is usually already there and the turn does not wait on the memory server at all. The first turn of a conversation has nothing to read ahead of, so it waits on the server normally (up to autoContextTimeoutMs). Set prefetch: false to query the message you just sent instead — at the cost of that wait on every turn. The query is not just that one message: with the default recallContextTurns: 5, the last few turns of the conversation ride along under a Prior context: block (one line per message, capped at 1000 characters), so a memory about the topic you were discussing a turn or two back can match — set it to 1 for the plain single-message query. When a turn's recall matches nothing new, you get a "no new memories this turn" row instead of a repeated block: it names the memories still in effect, one line each, so you can see exactly what was applied without the whole block repeating.

Giving the plugin its key

apiKeyRef: HINDSIGHT_API_KEY is only a name — the plugin looks up the value by that name on every call, in this order: the environment, then ~/.dsh/.credentials.yaml, then .env files. The usual way is one line in ~/.dsh/.credentials.yaml, added alongside your existing entries (like the LLM key — don't replace them):

version: 1
refs:
  HINDSIGHT_API_KEY: <the Hindsight server's token>