Agents that Remember — Introducing Cloudflare Agent Memory (Apr 17, 2026)

Summary

Cloudflare launched a private beta of Agent Memory — a managed service that extracts, classifies, deduplicates, and retrieves knowledge from agent conversations. The architecture solves "context rot" via compaction-triggered ingestion and a multi-channel retrieval pipeline with RRF fusion. The most transferable ideas are the 4-type memory taxonomy, topic-key supersession, and the HyDE retrieval technique. Our read: this is the first managed memory layer we've seen that treats staleness as a first-class design constraint — most agent memory projects stop at retrieval and hope the index stays clean on its own.

Key Points

▸ 1. The 4-Type Memory Taxonomy

What: Cloudflare classifies every extracted memory into exactly one of: Facts (what is true right now), Events (what happened at a time), Instructions (how to do something), Tasks (what is being worked on — ephemeral). Why it matters: Tasks are excluded from vector index entirely; facts supersede rather than accumulate. This prevents stale data from polluting retrieval and keeps the index lean. Apply to:

•vybeclaw / VybeMind (): Your current schema has concepts + patterns. That's 2 types vs 4. The missing types are events (e.g., "deployed X on Apr 28") and tasks (in-flight work). Adding them to the Zod schema in would let the drain pipeline prune stale tasks automatically.

•vybecoding / AgentSin Bureau (): Persona memory likely conflates facts and instructions. Separating them would enable targeted supersession — when a persona learns a new workflow, old instructions for the same slug get marked superseded rather than duplicated.

2. Topic-Key Supersession (Atomic Fact Versioning)

What: Each Fact and Instruction gets a normalized topic key. When a new memory has the same key, the old memory gets a forward pointer and is retired — not deleted. A version chain forms. Why it matters: Without this, every ingest creates new entries and old ones linger. The wiki grows unbounded and stale facts pollute recall. This is the mechanism that makes memory useful over months, not just sessions. Apply to:

•vybeclaw (, ): VybeMind currently does "action": "create | update" — but the update check is fuzzy (just matching slugs). A deterministic topic key (e.g., SHA-256(slug + type)[:16]) would make supersession exact. The drain job at is the right place to enforce the forward-pointer chain.

•vybecoding MEMORY.md (~/): The manual memory index has no versioning. Stale entries (e.g., the Oracle provision note) pile up. The /auto-memory skill could emit a supersedes: field so memory-trim knows which entries are dead.

3. Compaction as the Ingestion Trigger (vs. Stop Hook)

What: Cloudflare fires ingestion when the harness compacts context, not on session end. Compaction happens mid-session when context approaches model limits — so memory is captured even if a session runs for hours without stopping. Why it matters: VybeMind fires on Claude Code's Stop hook. Long sessions that exceed the context window and get auto-compacted mid-run lose knowledge that was pruned before the Stop event fires. Worth noting: we hit this exact failure mode during a long Studio generation run — the Stop hook fired on a truncated transcript, and the context with the most signal had already been discarded. The PreCompact hook fix is the most urgent item on this list. Apply to:

•vybeclaw (): Claude Code emits a PreCompact hook alongside Stop. Add VybeMind as a PreCompact hook handler (in under hooks.PreCompact) so compacted content is ingested before it's discarded — not after the session ends. This is a direct gap today.

4. HyDE Retrieval (Hypothetical Document Embedding)

What: Instead of embedding the query ("What package manager does the user prefer?"), also embed a hypothetical answer ("The user prefers pnpm over npm") and search with both vectors. The answer vocabulary matches stored memories better than the question vocabulary. Why it matters: Standard vector search misses on abstract or multi-hop queries because the question and the stored fact use different words. HyDE closes that gap significantly. Apply to:

•vybeclaw / VybeMind recall (): If/when you add vector search to VybeMind recall, generate a HyDE statement via a cheap Haiku call before embedding. Costs ~50 tokens but measurably improves recall on "how do we X" queries.

•vybecoding AI Search (): Any future semantic search over the wiki or agent memory should add a HyDE pass. The existing aiGateway.ts can produce the hypothetical statement cheaply.

5. Constrained Agent Tool Surface (Anti-pattern: Raw DB Access)

What: Cloudflare deliberately prevents agents from querying the raw memory store. The tool surface is intentionally limited to ingest / remember / recall / list / forget. Agents that can design their own storage queries "burn tokens on storage strategy instead of the actual task." Why it matters: This is an architectural principle, not just an API choice. Agents with raw filesystem or DB access waste context and produce unpredictable memory hygiene. Apply to:

•vybeclaw bots: Bots like trading-strategist.js that may read raw files for context should route through a narrow recall interface rather than free-form fs.readFileSync on wiki files.

•vybecoding AgentSin (): If personas can query the DB directly for their own memory, consider wrapping that in a narrow recallPersonaFact(slug, query) function rather than open ctx.db.query — both for token efficiency and to enforce the supersession chain.

6. Idempotent Ingestion via Content-Addressed IDs

What: Each message gets a SHA-256 hash of (sessionId + role + content) truncated to 128 bits. Re-ingesting the same conversation is a no-op. Why it matters: Safe to re-run, safe to retry, no duplicate knowledge accumulation. Apply to:

•vybeclaw (): VybeMind uses a per-CWD debounce stamp (STAMP_DIR`) to avoid re-ingestion. That's time-based, not content-based — two different transcripts from the same CWD within 10 minutes can't both be ingested. Replacing the stamp with a content hash of the transcript path + last-line would be more precise and eliminate both the false-negative (blocked legit ingest) and false-positive (ingests same content twice after a restart).

Action Items (Prioritized)

We're treating §3 (PreCompact hook) and §2 (topic-key supersession) as week-one priorities — both address failure modes we've already hit in production rather than theoretical improvements.

Source: blog.cloudflare.com