qdrant.yaml 5.6 KB

12345678910111213141516171819202122232425262728
  1. name: qdrant
  2. repo: https://github.com/kuehnberger/hermes-qdrant-memory
  3. sha: 5241e51274c657251776659bfb651aae4ba4bf1d
  4. version: "0.1.8"
  5. image: https://raw.githubusercontent.com/kuehnberger/hermes-qdrant-memory/5241e51274c657251776659bfb651aae4ba4bf1d/docs/banner.jpg
  6. description: "Dense semantic memory for Hermes, stored in a Qdrant collection you host. Each turn becomes one point — user and assistant text verbatim, no LLM summarization step — embedded on-device at 384 dimensions with fastembed (ONNX; the default install is 206 MB of runtime dependencies with no torch, versus ~1.2 GB when sentence-transformers was a hard dependency) and recalled by cosine search with an optional session_id payload filter, so a result quotes what was actually said. Registers 7 tools: qdrant_search, qdrant_upsert, qdrant_recall, qdrant_collect, qdrant_prepare, qdrant_forget (point-targeted deletion, dry-run by default, no delete-all), and md_search (local markdown knowledge index). An optional recall-indicator display, display.level (off | summary | verbose, default off), adds latency and collection to the chat indicator at summary and one numeric line per operation to the log at verbose — numbers and identifiers only, never recalled message text or search-result snippets (the md_search line also echoes the query you searched — your own input, not recalled content); it is separate from the older top-level progress knob (off | minimal | verbose), which gates status events. sentence-transformers is selectable on GPU hosts; both backends produce identical vectors for all-MiniLM-L6-v2, so switching backends needs no re-embedding. A circuit breaker (5 consecutive I/O failures, 120 s cooldown) degrades to no new memories instead of crashing, and is_available()/check_backend()/unavailable_reason() distinguish a misconfiguration from a dead server. Optional hybrid dense+sparse RRF search (default off): one config key enables dense+BM25 RRF fusion. A fresh collection declares the sparse vector at creation and an existing dense-only store gets it added in place (Qdrant >= 1.18) with one bundled backfill command; measured +0.024 recall@10 on the LongMemEval 1,000-example benchmark. INT8 quantization is not implemented. Local embedded Qdrant mode is not supported: the provider is REST-only and needs a reachable server. Disclosure — all memory content (verbatim user and assistant turns) is stored in and read back from the Qdrant server at the configured url (default http://localhost:6333); QDRANT_API_KEY is only needed for Qdrant Cloud, is kept in .env and read through the Hermes secret scope, and is never written to config.yaml or qdrant.json. Disclosure — the only other outbound traffic is a one-time download of the embedding model weights (plus the 10 MB Qdrant/bm25 sparse model when hybrid is on) from huggingface.co (fastembed falls back to storage.googleapis.com/qdrant-fastembed if that fails); embedding runs on-device and no memory content goes to any other party. Disclosure — settings are stored in <HERMES_HOME>/qdrant.json (created 0600, no credentials) and last-operation status in <HERMES_HOME>/qdrant-status.json. Disclosure — md_search returns snippets of local markdown to the model from ~/.hermes/skills, the Obsidian vault at ~/Documents/hermes_default, ~/.hermes/hermes-agent/website and any path in the md_docs_roots key of qdrant.json; its index is built only when you run scripts/md_ingest.py and lives in <base Hermes home>/state/md-search/ (shared by all profiles). Disclosure — the optional semantic tier (md_ingest.py --semantic, and md_search queries with few lexical hits) sends embedding vectors plus file-path labels and section headings, never chunk text, to a separate hermes_md_docs collection on the Qdrant server named by the QDRANT_URL env var (default http://localhost:6333, no API key), and downloads paraphrase-multilingual-MiniLM-L12-v2 from huggingface.co on first use. Disclosure — the plugin is not a server and registers no MCP server, but it is loaded in-process into every Hermes process that embeds, so each one pays its own model copy (measured ~194-211 MB per process). Disclosure — the bundled scripts/backfill_memory.py (run by hand; a dry run unless --yes) reads a profile's Hermes session database (<profile>/state.db, read-only) and copies every stored user, assistant and tool message from every session source (cli, cron, subagent, gateway), in 2,000-character chunks, into the collection at the url in qdrant.json (default http://localhost:6333); nothing reads state.db unless you run it."
  7. maintainer: kuehnberger
  8. tier: community
  9. category: memory
  10. requires_hermes: ">=0.21.5"
  11. docs_url: https://github.com/kuehnberger/hermes-qdrant-memory#readme
  12. known_issues:
  13. - "Windows on ARM is unsupported: grpcio, pulled in by qdrant-client, has no win_arm64 wheel. Linux and macOS on x86_64/arm64 and Windows x86_64 are supported."
  14. - "Local embedded Qdrant mode is not supported — the provider is REST-only and requires a reachable server."
  15. - "md_search returns 'no markdown index yet' until you run the bundled ingest script once (python scripts/md_ingest.py, lexical only and fast; add --semantic for the embedding tier, which takes about 7 minutes for 5500 chunks on a 5-core CPU). No background process does this for you."
  16. - "Running the plugin's pytest suite inside an installed copy still changes the PM dependency stamp (the .pytest_cache directory is hashed as a build input); only an upstream core change can exclude gitignored state."
  17. capabilities:
  18. provides_tools:
  19. - qdrant_search
  20. - qdrant_upsert
  21. - qdrant_recall
  22. - qdrant_collect
  23. - qdrant_prepare
  24. - qdrant_forget
  25. - md_search
  26. provides_hooks: []
  27. provides_middleware: []
  28. requires_env: []