Memory browser (reflect serve)

reflect serve is a local web utility for browsing, searching, and curating the reflect knowledge base. It ships with the CLI documented on the reflect CLI reference; for the engine’s recall pipeline see Recall reference.

reflect serve                       # http://127.0.0.1:8377, KB at $GLOBAL_LEARNINGS_PATH or ~/.learnings
reflect serve --port 8942           # pick a port
reflect serve --repo /path/to/kb    # browse a specific KB

To reach it from another device on your tailnet, front it with a proxy — the server itself never leaves loopback:

tailscale serve --bg --https 8942 http://127.0.0.1:8942

What it does

• Browse — a faceted memory list (confidence · type · scope · tags/projects), sortable by newest, computed recall score, or title.
• Search — lexical BM25 ranking; each result carries a match score and a browse-ordering score (confidence × recency × tag overlap). This is a browse heuristic, not the recall reranker’s score — the real reranker (see R2/R3 in the recall reference) deliberately dropped exp-decay recency. Full semantic search stays on the reflect search CLI.
• Graph — a two-layer force graph of memory and entity nodes; edge widths encode graphml relation weights. Toggle the entity layer, drag/pan/zoom, click a memory to open it.
• Timeline — memories grouped chronologically.
• Superseded — superseded_by chains rendered newest tip to oldest ancestor.
• Curate — soft-archive/restore (reversible, no hard delete), edit a note’s confidence, and multi-select memories to queue a group for compression by the /reflect consolidate skill (written to a versioned compress-queue.yaml — the web app never calls an LLM).
• Stats — confidence/type/scope/tag distributions and engine-op usage counts.

Light and dark themes throughout.

Weightages surfaced

Weightage	Where
Editable confidence	segmented control in the detail drawer
Computed browse score	on every card and in the drawer
Graph edge weights	edge width in the Graph view
Recall usage stats	engine-op counts in the Stats view

Curation is metadata-only and file-first

Mutations edit the markdown note (the source of truth) and the browser view reloads immediately. The nano-graphrag cache used by reflect search is not rebuilt synchronously — the engine only supports a full-batch reindex — so after archiving or editing confidence, run reflect reindex to refresh semantic search and the entity graph. The UI hints at this. Note bodies stay agent-authored; only metadata is editable here. Postgres backends are read-only in this release.

Soft-archive moves the note into an archived/ sibling directory. This is the browser’s own reversible delete and is separate from the forget sweep’s .forgotten/ directory (see A3 in the recall reference), which the sweep manages with its own database accounting. Archiving here does not run the sweep or touch its records.

Loopback only — no authentication

Curation is guarded: the server rejects any request whose Host isn’t loopback (defeats DNS-rebinding) and requires an X-Reflect header on mutations, so only the same-origin SPA can drive them. There is no auth — do not bind a public interface. To reach it across your tailnet, use the tailscale serve proxy shown above rather than exposing the port directly.

Screenshots

Source

src/reflect_kb/serve.py and the e2e suite under tests/e2e/ in stevengonsalvez/ainb-reflect-memory. Full guide: docs/reflect-serve.md.