Introduction
Recall is memory infrastructure for AI apps that combines semantic search with record-based storage — so AI can actually run computations on what it remembers.
The Problem
Section titled “The Problem”Existing memory solutions (graph + vector databases) handle relationships and fuzzy recall well, but you can’t compute on them. Ask “how much have I spent on groceries this month?” and the LLM will hallucinate a number.
Recall adds structured, record-based memory alongside semantic memory. This unlocks precise aggregations — sums, averages, filters, counts — that LLMs otherwise get wrong.
Two Memory Systems
Section titled “Two Memory Systems”| Semantic Memory | Structured Memory | |
|---|---|---|
| Best for | Facts, preferences, context | Trackable data, logs, records |
| Storage | Embeddings + vector search | Schema-defined tables |
| Retrieval | ”What does the user like?" | "Sum all payments to Jayden” |
| Example | ”User prefers dark mode” | { recipient: "Jayden", amount: 150 } |
Both systems run in your existing database. Zero vendor lock-in.
Why Recall?
Section titled “Why Recall?”Building AI memory is harder than it looks:
- Extraction — What facts should be remembered from a conversation?
- Deduplication — How do you avoid storing “User likes coffee” 50 times?
- Retrieval — How do you find relevant memories without loading everything?
- Computation — How do you answer “how much?” or “how many?” accurately?
- Infrastructure — Where do you store this? Another service to manage?
Recall handles all of this. Semantic memory for fuzzy recall, structured memory for precise queries — both running in your existing database.
Key Features
Section titled “Key Features”- Two memory systems — Semantic for fuzzy recall, structured for precise computation
- LLM-powered extraction — Automatically identify facts from conversations
- Smart consolidation — Deduplicate with ADD/UPDATE/DELETE decisions
- Schema-based storage — Define Zod schemas, get automatic SQL queries
- Pluggable architecture — Swap databases, embeddings, and extractors
- Zero lock-in — All data stays in your existing infrastructure
Packages
Section titled “Packages”Recall is modular. Combine packages based on your stack:
| Category | Available |
|---|---|
| Core | @youcraft/recall (semantic memory) |
| Structured | @youcraft/recall-structured (record-based memory) |
| Database Adapters | SQLite, PostgreSQL, MySQL |
| Embeddings | OpenAI, Cohere |
| Extractors | OpenAI, Anthropic |
| Integrations | Vercel AI SDK, MCP Server, MCP Tools |
See Packages for installation and details.
Get Started
Section titled “Get Started”- Overview — See how Recall works with code examples
- Quickstart — Build a memory-enabled chatbot in 5 minutes
- Core Concepts — Learn how AI memory systems work
- Structured Memory — Schema-based memory with precise queries