Frequently asked.
Common questions about agent memory architecture and Recall — concepts, deployment, pricing, and security.
Concepts
- What is agent memory?
- Agent memory is durable, structured state — facts, preferences, events, entities, and relations — that an LLM agent persists outside its context window and retrieves on demand. It is what makes an agent feel continuous across turns, sessions, and tasks.
- How is agent memory different from RAG?
- RAG retrieves from a static document corpus. Agent memory writes new state from interactions, types it, ages it, supersedes it when contradicted, and reasons over relations. RAG is a read-only library; memory is a living workspace.
- How is agent memory different from a vector database?
- A vector database stores embeddings and returns nearest neighbors — useful storage substrate. A memory system adds typed schemas, multi-retriever fusion, supersession, decay, drift detection, and confidence scoring on top. Calling a vector DB 'memory' is like calling Postgres 'an application.'
- When should I use agent memory vs long context windows?
- Long context wins when the relevant material is small, fits in the budget, and changes per request (one-shot analysis of a document). Memory wins when state must accumulate from interactions and persist across sessions. Stanford's Lost-in-the-Middle work showed long context degrades on large prompts; memory + retrieval avoids that.
- What memory types are essential?
- Five canonical types: fact (stable predicate), preference (mutable choice), event (timestamped occurrence), entity (identity), relation (typed edge). Treating memory as flat text collapses these and loses temporal and relational signal.
About Recall
- What is Recall?
- Recall is an open-source agent memory layer built in Rust. It implements a 7-stage write pipeline, typed memory schema (fact / preference / event / entity / relation), 5-retriever hybrid retrieval with RRF fusion, freshness decay, supersession, drift detection, and 3-layer hallucination defense.
- Is Recall open source?
- Yes — Apache 2.0. The Rust core, TypeScript and Python SDKs, visualization library, and architectural spec are all open. github.com/arc-labs.
- What's the difference between Recall and Mem0?
- Mem0 is the easiest agent memory framework to start with — three lines of Python from zero to working. Recall is more opinionated about quality at write time: 7-stage filtering pipeline, typed schemas, multi-retriever fusion, drift detection. Both are open source. Recall fits long-running production conversational agents; Mem0 fits prototypes and demos.
- What's the difference between Recall and LangChain Memory?
- LangChain Memory abstractions (BufferMemory, WindowMemory, SummaryMemory) are conversation-scoped — they handle continuity within a single session. Recall is durable, cross-session memory with types, supersession, and decay. They compose: LangChain for in-session, Recall for cross-session.
- Does Recall replace my CRM / knowledge base / vector database?
- No. Recall complements them. The CRM / KB stays the system of record; Recall is the working surface that makes their data usable in real-time agent prompts. The vector DB sits underneath Recall as the storage substrate (pgvector by default).
Deployment & hosting
- How do I run Recall?
- Three modes: embedded (in-process SQLite + sqlite-vec — perfect for prototypes, scales to 100K memories per user), self-hosted (you operate Postgres + pgvector), or managed cloud (we operate the infrastructure). Switch modes with one config line.
- What languages and runtimes does Recall support?
- Rust core; first-party SDKs for TypeScript (via napi-rs) and Python (via PyO3). REST API for everything else. Node 20+, Python 3.10+, Rust 1.78+.
- What database does Recall use?
- By default: SQLite-vec for embedded mode, Postgres + pgvector for self-hosted and managed cloud. Recall does not require a separate vector database — though you can wire one in if you want.
- Can I self-host the managed cloud version?
- Yes — the open-source Recall is the same software the managed cloud runs. Self-hosted deployments use Postgres + pgvector and the Rust core; we provide deployment guides for AWS, GCP, and bare-metal.
- What scale does Recall support?
- Embedded mode is comfortable up to ~100K memories per user. Single-Postgres mode handles 10M memories with HNSW. Sharded Postgres goes to 100M. For 1B+, pair with a specialized vector store like Qdrant.
Pricing
- How much does Recall cost?
- The open-source Recall is free under Apache 2.0. The managed cloud is in design-partner preview; pricing is per-tenant rather than per-vector. Contact sales@arc-labs.ai for current pricing.
- Is there a free tier?
- The open-source version is free, full-featured, and ships with the same core as the managed cloud. The managed cloud preview is currently invite-only.
Security & compliance
- What's Recall's security posture?
- TLS 1.3 in transit, AES-256-GCM at rest. Per-tenant isolation at row and query level. Per-tenant data keys via envelope encryption. SOC 2 Type II audit in progress (target Q1 2027). GDPR-compatible deletion. See /security for full details.
- Is Recall HIPAA compliant?
- Not currently. Contact us if your deployment requires HIPAA — we can scope a path forward.
- How does Recall handle deletion requests?
- Hard-delete by default. Memories are removed from primary storage and all secondary indexes (vector, lexical, graph) within 24 hours. The audit ledger retains a tombstone (memory ID, deletion timestamp, requester) but not memory content.
- Where does Recall store data?
- For self-hosted: wherever your Postgres lives. For managed cloud: US-East, EU-West, or AP-South — selectable at create time. Cross-region data movement is opt-in.
Still have questions?
The Learn track has 28 in-depth pages, or write to us directly.