Tallei

Ghost memory layer for ChatGPT, Claude, and Gemini. One place, all your AIs.

Tell one AI how you like things. Every other AI already knows. No re-explaining, no copy-pasting, no starting from scratch.

---

Recall: The Journey

Apr 14  ████████████████████████████  ~30,000ms  baseline
Apr 15  ██████████████████████████    ~24,000ms  in-memory cache
Apr 16  ████                          ~717ms     multi-layer cache — WRONG on first call
Apr 20  ██                            ~30ms / ~400ms  three-bucket recall, always correct

Date	p50	Approach	Accurate?
Apr 14	~30s	Full pipeline sync per call	✓
Apr 15	~24s	In-memory cache (60s TTL)	✓
Apr 16	~717ms	Multi-layer + async enrichment	✗ first call
Apr 20	~30ms / ~400ms	Three-bucket — always sync	✓ always

Apr 16 looked like a win. It wasn't. Fast wrong answer is worse than slow right answer.

---

Three-Bucket Recall

recall_memories(query)
        │
        ▼
  LRU in-process cache  ~0ms  (10 min)
        │ miss
        ▼
  Redis exact cache      ~5ms  (120s, query-hash keyed)
        │ miss
        ▼
  Postgres: single fetch, decrypt all rows
        │
        ├── PREFERENCE bucket  (1,500 tok)  pinned → refcount order
        ├── LONG-TERM bucket   (4,800 tok)  dump-all ~30ms | hybrid ~400ms
        │     overflow: vector top-25 + BM25 top-15 + temporal top-5
        └── SHORT-TERM bucket  (1,700 tok)  ≤60 days, recency-first
        │
        └── write to LRU + Redis → return

Token estimation: ceil(chars / 4) — no tokeniser, zero overhead.

Sorting within buckets: preferences sort by pin → frequency; long-term dump-all by frequency; long-term overflow by a combined vector + BM25 + activity score; short-term strictly by recency with a decay penalty. See ADR-011 for the exact formulas.

What's coming:

• Promotion — short-term memories referenced 3+ times graduate to long-term automatically
• Eviction — long-term memories untouched for a year get soft-archived (still searchable, excluded from dump-all)
• Per-user short-term window — currently 60 days hardcoded, will tune based on memory volume

Deliberately not doing: LLM rerank (the host model already does this in-context for free), HyDE pre-embedding (adds latency, unnecessary for most users).

---

Memory System

Types

Seven types, classified automatically at save time.

Type	Bucket	Detected by
preference	preference	always pinned; identity, settings, favorites
fact	long-term	default
decision	long-term	architectural choices, commitments
event	short-term	timestamped occurrences
note	short-term	loose observations
lesson	long-term	"best practice", "learned that", "never do"
failure	long-term	"broke", "crashed", "outage", "postmortem"

lesson and failure are first-class — they land in long-term and get recalled like facts, not buried.

Access Freshness

Every recall hit bumps reference_count and updates last_referenced_at. Overflow scoring weights frequently-accessed memories higher, with exponential decay over 21 days so stale memories fade naturally.

Provenance & Confidence

Confidence is inferred from reference_count — UNCONFIRMED at 1, MED at 2+, HIGH at 5+. Every memory also carries a platform tag and a provenance block (platform + timestamp) so you always know where it came from. Cross-platform conflicts on the same subject get flagged. Preferences supersede each other — old version soft-deleted, latest canonical.

---

Fire-and-Forget Saves

save_memory() → return ~15ms ✓
               [background] summarize → embed → upsert vector → dedup check

Never await the full pipeline in the MCP handler. Background worker handles the rest.

---

Gotchas

• Fire-and-forget is intentional. Don't await the save pipeline in the MCP handler.
• Cache everything that hits OpenAI or pgvector. OAuth token cache (10 min), LRU recall (10 min), Redis (120s).

---

Docs

• Setup guide
• Deployment
• ADR-011: Three-bucket recall
• ADR-012: Graph layer removal

---

Business Source License 1.1.