dsh-hindsight-manager

English · 中文

Manage the local Hindsight memory daemon from inside DeepSeek Harness: lifecycle, guided setup, and a page to browse and search the memory bank.

Features

• Starts with DSH, stops with DSH: retries and a watchdog, so no daemon is left behind.
• Guided setup: install the integration → add a model key → pick a cache dir → start; only the missing step is shown.
• Bank overview: documents / facts / links / observations, semantic search, recent documents.
• Memory graph: entity co-occurrence graph; focused on the busiest entity by default, full graph on demand, click a node for its memories.
• Growth replay: replays the graph by memory write time, with a scrubber and play/pause.
• Bank health: last write/consolidation, pending and failed operations, plus a timed recall probe.
• Bank settings: recall token cap, budget strategy, retrieval tier, rerank, observations, auto consolidation, extraction mode.
• Cleanup: near-duplicates / untouched / observations / forgotten lists; bulk invalidation with two confirmations (default batch limit 100, adjustable); duplicate groups keep one copy by default.
• Observation history: click observation in the graph drawer or an observation row in cleanup to see versions written by consolidation and their sources.
• Bank switching: choose which bank to read (read-only; writes still follow the session workspace).
• Status chip in the sidebar; no console window; dead-proxy bypass.

Install

git clone https://github.com/fffzzzjjj/dsh-hindsight-manager.git
dsh plugin --profile <your profile> add file:<cloned dir>

Restart DSH; "Local Memory" appears in the sidebar.

Update

There is no auto-update: dsh plugin simply forwards to pnpm, and a file: install has to be re-added:

cd <cloned dir>
git pull
dsh plugin --profile <profile> remove dsh-hindsight-manager
dsh plugin --profile <profile> add file:<cloned dir>

Then restart DSH. Diagnostics shows the current version and a Check for updates button that compares against GitHub and prints these commands.

For client-only (client.js) releases, reinstalling plus a page refresh (Ctrl+R) is enough; index.js changes require a DSH restart.

First run

Open "Local Memory" and follow the four steps:

1. Install the memory integration (@vectorize-io/hindsight-coding-agents), then restart DSH so the hindsight_* tools appear.
2. Add the model key — stored only in ~/.hindsight/profiles/coding-agent.env, never echoed back.
3. Choose a cache directory (optional) — an absolute path or the Browse… picker.
4. Start the daemon — the first start downloads local models (a few minutes); later starts take 20–40 s.

After that the checklist collapses into "Ready", and conversations are written to memory when they end.

Pages

• Overview: stats, search, recent documents, settings, bank health, bank settings, diagnostics.
• Graph: focused on the busiest entity (two hops); click to refocus, blank area or × to return to the full graph; custom "min mentions"; growth replay; the drawer lists that entity's memories.
• Cleanup: four lists and bulk invalidation; restore from "Forgotten".
• Usage: token spend of the extraction/consolidation calls — today and for 24 h / 7 / 30 days, a bar chart with fixed-width bars (by hour for the 24-hour range, by day otherwise), split by operation and by model, plus the latest calls. Read from the daemon's llm-requests log; local embeddings and rerank cost nothing and are excluded. Cost is estimated from DeepSeek's published prices with the peak/idle difference applied per call — a calculation, not a bill.
• Retrieval quality: which embedding/rerank models are actually in use, the vector size, how much of the bank is Chinese, and a timed recall that reports latency, result count, score spread and the top hits. Warns when an English-only encoder is indexing a Chinese bank, or when rerank scores are too bunched to order anything.
• Recall self-test: your own questions, scored as recall@1/@3/@5 with latency. Save a run as a baseline and the next run reports the delta — so a tuning change is measured, not guessed. Questions are never generated from the bank: a question written out of the memory it is meant to retrieve would score the bank against itself.
• Settings: cache dir, start with DSH, footer status, sidebar entry (also on Plugins → Installed → Local Memory).

Config

<DSH_HOME>/storages/dsh-hindsight-manager/config.json:

Field	Meaning
installDir	cache/model directory; empty = system default
autoStart	start the daemon with DSH
showFooterStatus / showSidebarEntry	footer chip / sidebar entry
bank	pin a bank to read; empty = automatic
searchBudget	tier for the panel's search and probe: low / mid / high

FAQ

Symptom	Fix
"Start failed" with Connection error / WinError 1225	The system proxy points at a closed port. The plugin bypasses it automatically; if it still fails, restore the VPN/proxy and retry.
"Starting" for over 6 minutes	First-run model download; check ~/.hindsight/profiles/coding-agent.log; failures retry with backoff.
Port 9077 in use	netstat -ano | findstr :9077, then stop that process.
No "Local Memory" entry	Restart DSH after install/update.
No hindsight_* tools in a session	The integration is installed but DSH was not restarted.
Slow search	Set the retrieval tier to low, or turn reranking off (faster, less relevant).

Limits

• Windows only: process management uses netstat / taskkill.
• Search uses recall only (no reflect); deleting or exporting documents is not supported yet.

License

MIT. The watchdog logic is adapted from Hua1Q1nG/dsh-hindsight-local-optimize; the memory engine is Vectorize Hindsight.