Core Concepts

Core Concepts

How Awareness thinks about memory — and why it's different from just storing text.


Memory Space

A Memory in Awareness is a project-level container — like a dedicated workspace for one team, project, or AI application. Everything stored inside stays together: conversations, decisions, team preferences, and knowledge history.

You can create one memory per project, or share a memory across multiple AI agents working on the same codebase.


Events: The Raw Feed

Events are the raw material Awareness processes. Anything worth remembering is an event:

  • A conversation between you and your AI assistant
  • A tool call result (e.g., ran tests, deployment succeeded)
  • A code change with a description
  • A design decision made during a session

Events are lightweight to record. You don't need to organize them — Awareness does that automatically.


Knowledge Cards: Where Understanding Lives

Knowledge Cards are what Awareness creates from events. If an event is a raw conversation, a knowledge card is the insight extracted from it.

Think of it this way: a conversation where you say "we tried Redis for session storage but it caused memory spikes under load" becomes a knowledge card of type pitfall with the summary: "Redis session storage causes memory spikes at scale — prefer Postgres-backed sessions."

Awareness extracts 13 types of knowledge cards:

TypeWhat it capturesExample
skillRepeatable steps (muscle memory)"Deploy: pull → build → nginx reload"
decisionArchitecture or product choices"Auth uses RS256 — HS256 rejected due to key rotation complexity"
problem_solutionBug fixes and diagnosis"502 error after deploy = nginx not reloaded"
pitfallRisk warnings and known traps"Never run db push in production — drops columns"
insightHigh-level patterns worth repeating"Feature flags reduce rollback risk for DB migrations"
workflowFirst-time process recordings"How we set up the CI pipeline"
key_pointImportant facts"Backend stack: FastAPI + PostgreSQL"
personal_preferenceIndividual coding habits"Always use snake_case; TypeScript strict mode required"
activity_preferenceWork style habits"No major refactors on Fridays"
important_detailEasy-to-miss context"API keys use the prefix aw_"
plan_intentionFuture plans"Planning to migrate auth to OAuth2 next sprint"
health_infoPersonal context"Prefers dark mode; sensitive to bright screens"
career_infoBackground context"Full-stack engineer, Python/Go primary"

Conflict Detection

When a new card contradicts an existing one, Awareness automatically detects it. The old card is marked superseded, and the new card is linked to it — so you can always trace the evolution of a decision or fact.


Intelligent Recall

When you ask Awareness a question, it doesn't just do a text search. It combines:

  • Semantic matching — finding conceptually related knowledge, not just keyword hits
  • Freshness scoring — recent, actively used cards rank higher
  • Knowledge graph traversal — following relationships between connected cards

The result is relevant, not just recent. You get what matters for this situation, not everything ever stored.

Recall Modes

ModeBest for
hybridDefault. Combined vector + keyword + structured search for best overall results
autoAutomatically selects the best retrieval strategy based on your query
preciseVector search only — targeted semantic matching for exact fact lookup
sessionFull session history — expands matched chunks to complete session context
structuredDB-only fast lookup — zero vector search, pure structured data (<50ms)

Progressive Disclosure

For broad queries, Awareness first returns summaries. If you need the full detail, it fetches it. This means your AI agent reads only what's necessary — saving tokens without losing depth.


Multi-Agent Collaboration

Awareness is built for teams of agents, not just one.

Each agent can have a role — a named persona like builder_agent, reviewer_agent, or devops_agent. Roles enable cross-agent recall: a reviewer can see what the builder learned. A new agent joining the project inherits the full team knowledge instantly.

Data Isolation

Multiple users (or agents) can work in the same memory with isolated views:

  • Each user's private data is tagged to them
  • Shared team knowledge is visible to all agents
  • Fine-grained read/write permission control is available

How Knowledge Evolves Over Time

Awareness isn't a static archive — it's a living knowledge base. Here's how memory evolves:

  1. Record — Events arrive from conversations or tool calls
  2. Extract — Knowledge cards are created automatically
  3. Deduplicate — Near-identical cards are merged; stale cards are flagged
  4. Conflict check — Contradictions are detected and resolved
  5. Cluster — Related cards are grouped by topic for efficient broad recall
  6. Score — Cards are ranked by relevance, recency, and usage signals
  7. Recall — The right knowledge surfaces at the right moment

Integration: MCP or SDK

MCP (Model Context Protocol) — Best for IDE agents (Cursor, VS Code, Windsurf, Claude Code). Zero configuration. Four unified tools: init, recall, lookup, record.

SDK (Python / TypeScript) — Best for custom agents and framework integrations (LangChain, CrewAI, AutoGen). Full programmatic access, batch operations, file uploads.

Both methods access the same underlying knowledge base.


Next Steps