Memory Graph

Visual audit surface + memory API for Friday, a 24/7 AI assistant built on Claude Code.

Flask + SQLite backend, single-file D3.js frontend. No external DB, no vector store, no framework — just one Python file and one HTML file.

⚠️ This repo is not a template to clone. It is a snapshot of my personal instance, committed here for reference and as the source for the screenshots in the showcase. The code in this repo is generated autonomously by Claude Code when it runs the SETUP.md from friday-showcase — every user ends up with their own slightly different version. If you want to stand up your own Friday, follow the SETUP from friday-showcase; don't clone this one.

Version: 2.5.1 — see VERSION constant in api_server.py.

Tabs

Six tabs served at /graph:

• Graph — force-directed D3.js visualization of conversations, memories and entities as interactive color-coded nodes (drag / zoom / filter).
• Logs — chronological view of conversation logs with collapsible date groups and live search.
• Arch — system architecture diagram with draggable nodes. Positions persist server-side via /kv/<key>.
• RAG — semantic search dashboard: hybrid search (FTS5 + cosine RRF), embedding stats, reindex trigger.
• Brain — consolidated audit surface for Friday's self-improving harness (Goals, Plans, three-layer Memory, causal World Model, Capabilities + Autonomy, Verifier + Sandbox, Experiments, Metrics). Sticky sub-nav, responsive grid. See the harness deep-dive.
• Crons — two-column diff: runtime-active jobs with live countdowns to next fire, vs the prompts persisted in ~/.claude/cron-prompts.md, with sincronizado / ⚠ no corriendo badges.

Screenshots

Graph

Logs

Architecture

RAG

Brain

Crons

Stack

• Frontend: Single-file HTML with D3.js v7 and vanilla JS.
• Backend: Flask + SQLite memory API (api_server.py) on port 7777.
• RAG: Gemini Embedding 001 (3072 dim), cosine similarity, RRF hybrid ranking, FTS5 for lexical.
• Self-improving harness (v2): 13 extra tables — goals, plan_tree, capabilities, autonomy_levels, wm_entities, wm_relations, wm_events, wm_predictions, verifications, sandbox_executions, experiments, metrics, active_crons — plus additive columns on existing ones (provenance, confidence, last_verified, layer, skill maturity, etc). All migrations are ALTER TABLE … IF NOT EXISTS style, so older databases upgrade in place.

Setup

cd ~/proyectos/memory-graph
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
FRIDAY_DB_PATH=~/.claude/memory.db python3 api_server.py

Open http://localhost:7777/graph.

API — core endpoints

The full API has grown past 100 endpoints. Core ones:

Endpoint	Description
GET /health / GET /version	Health + version
GET /graph	Serve the dashboard HTML
POST /conversation/log	Log a message {role, content, channel} with auto-importance scoring
GET /conversation/recent / search / stats	Read logs
POST /memory / GET /memory/list / /recall / /search	Long-term memories (typed: user / feedback / project / reference)
GET /memory/episodic / /semantic / /procedural	v2 three-layer views
POST /memory/<id>/verify / POST /memory/decay	v2 confidence decay + re-verify
POST /entity / GET /entity/search	Identity entities (who/what things ARE)
POST /wm/entity / /relation / /event / /prediction	v2 structured world model (state, S-P-O, causal events, testable predictions)
POST /goal / GET /goal/active / /next / PATCH /goal/<id>	v2 goal engine
POST /plan / POST /plan/<plan_id>/node	v2 hierarchical plan trees
POST /capability/<n>/record / GET /capability/can	v2 calibrated self-knowledge
POST /autonomy/check	v2 autonomy gate (L0 suggest → L5 self-modify)
POST /verify	v2 factual / consistency / hallucination checks
POST /sandbox/execute	v2 dry-run / simulation / live
POST /experiment / /observation / PATCH /conclude	v2 A/B experiments
POST /metric / GET /metric/summary	v2 KPI framework (11 known metrics)
GET /search/semantic / /search/hybrid	RAG
GET /embeddings/stats / POST /embeddings/reindex	Embeddings ops
GET /kv/<key> / PUT /kv/<key>	Key-value store (dashboard positions, etc)
POST /cron/active / GET /cron/active / GET /cron/prompts	Runtime cron snapshot + disk prompt parser

Full table ownership notes and the soft-observation → structured-knowledge promotion flow live in Friday's CLAUDE.md and in the harness deep-dive.

Architecture

The server is the single source of truth for everything Friday remembers, believes, plans or measures. Four broad concerns share one SQLite file:

• Conversations — raw message log, with importance classification, FTS5 index, and 3072-dim Gemini embeddings stored as BLOBs.
• Long-term memory — typed memories, entities, skills. Each row now carries provenance + confidence + last_verified.
• Self-improving harness — goals, plans, capabilities, autonomy, world model, verifier, sandbox, experiments, metrics. All additive.
• Runtime state — active cron snapshot, importance keywords, KV store.

Related

• friday-showcase — the 24/7 Claude Code assistant that writes to this memory server.
• harness.html — technical deep-dive on the v2 self-evolving harness.

---

Built by Bruno Silveira with Friday.