Changelog
What's new in takizen.
CoT Feedback System
AI agents can now learn which memory combinations work best in their reasoning processes. Memories are ranked by effectiveness over time, not just semantic similarity.
Added
-
cot_feedbacktool — record which memories were useful in a Chain of Thought session, with an effectiveness score (0–1) and optional notes -
recallnow acceptsboost_by_cotandcot_boost_weight— prioritize memories with a proven effectiveness history -
New Supabase tables:
cot_sessions,cot_session_memories,cot_feedback -
RPC functions:
match_memories_cot_boosted,decay_cot_feedback,get_cot_memory_score -
CoT decay at
0.02/day— more stable than memory decay, runs in the daily 2am cron
Frontend
- Landing page redesigned — dark-first palette, Supabase-inspired nav with beta badge
- Bento grid with 6 animated SVG feature cards (semantic memory, instant recall, strength decay, MCP native, memory graph, namespace isolation)
/docs— full MCP API reference/dashboard— rebuilt as structured table layout (Vercel/Cloudflare style) with overview-first stats, usage chart, and compact 7-column memory table- 404 page and
/healthendpoint
Compliance & Security
Legal infrastructure for GDPR compliance. All data operations are now auditable, consent is versioned, and users can submit deletion requests with a 30-day deadline.
Added
-
audit_log— append-only record of all operations (SQL rules block UPDATE/DELETE). Satisfies GDPR Art. 30 -
tos_acceptances— versioned ToS consent history per namespace. GDPR Art. 7 -
gdpr_requests— data subject requests (erasure, access, portability) with 30-day deadline tracking. GDPR Art. 12/15/16/17 -
key_prefixon API keys — first 10 chars stored in plaintext for visual identification without exposing the secret -
statstool — namespace usage counters (remember/recall this month, recalls today) - Monthly rate limit on
remember, configurable per plan
Changed
-
RLS enabled on all tables —
deny_allforanonandauthenticatedroles. Onlyservice_rolecan access from the Worker -
namespaces.api_key_hashremoved — API keys now live exclusively in theapi_keystable
Memory Graph
Memories are no longer isolated vectors — they form a graph. Related memories are linked automatically and surfaced together during recall via hybrid search.
Added
-
linktool — create, delete, or list typed relationships between memories. Types:related_to,supports,contradicts,part_of,caused_by -
Auto-linking on
remember— finds similar memories (threshold 0.75, max 5 links) and links them automatically. Best-effort: errors don't fail the operation -
Graph expansion in
recall— results include neighbor memories from the link graph - Hybrid search with Reciprocal Rank Fusion (RRF) — combines semantic similarity and graph traversal
-
updatetool — modify content and/or tags of an existing memory -
memory_linkstable with HNSW index - Full-text search — auto-generated
ftstsvector column onmemories
Initial Release
Core MCP server with four tools, semantic deduplication, automatic decay, and weekly consolidation.
Added
-
remember— store a memory with content, optional tags, and a 1536-dim vector embedding. Semantic deduplication at 0.92 similarity (updates instead of duplicating) recall— semantic search with optional tag filtersforget— archive to KV cold storage before deleting from Postgreslist— paginated list of namespace memories- Auth — SHA-256 hashed API keys, validated on every endpoint except
/health - Daily decay cron (2am) — reduces
strengthof inactive memories; archives at 0 - Weekly consolidation cron (Sunday 3am) — clusters similar memories using claude-haiku (Pro plan)
- HTTP/SSE transport via
@modelcontextprotocol/sdk
Stack
- Cloudflare Workers (TypeScript runtime)
- Supabase PostgreSQL + pgvector (VECTOR 1536, HNSW index)
- OpenRouter —
openai/text-embedding-3-smallfor embeddings,anthropic/claude-haiku-4-5for consolidation - Cloudflare KV for archived memory cold storage