Skip to main content
StremAI
Sign InSign Up

Memory model

How StremAI memory works.

StremAI is a shared memory layer for coding agents. It remembers useful work context, retrieves it for the next session, and gives owners controls for access, export, retention, and erasure.

The lifecycle

Store

Agents write patterns, pitfalls, decisions, test findings, dependencies, and project context into a project or agent scope. Writes include provenance metadata so later readers can understand where the memory came from.

Recall

Recall combines multiple search strategies, including tags, aliases, full-text search, and vector similarity. Results are ranked so the agent gets a compact set of likely useful memories instead of a raw database dump.

Age

Memory has health signals such as confidence, staleness, tier, and recall activity. Dreaming and repair jobs can surface stale, low-confidence, duplicate, or weakly linked entries for review.

Share

Project and team roles decide who can read, write, export, audit, or erase shared memory. Multiple agents can use the same project without becoming duplicate rows in the connection surface.

Govern

Exports, erasure requests, audit logs, revocation, retention, and project governance endpoints give owners a way to inspect and control memory after it has been written.

What gets remembered

The best memories are durable facts an agent should not have to rediscover: architectural decisions, recurring failures, setup pitfalls, dependency constraints, successful test fixes, and project conventions.

  • Connecting an agent does not automatically mirror your repository, terminal, browser, or local filesystem.
  • Project files and GitHub imports are explicit actions, not ambient background capture.
  • Soft delete hides memory from recall while preserving an audit trail; project hard delete is a separate owner-confirmed erasure flow.
  • Recall-event rows are retained for 30 days by default, while ordinary audit logs are pruned after 90 days with a retention audit row.
  • StremAI does not train AI models on customer memory content by default.

Need the practical boundary?

The companion guide explains what agents can and cannot remember in plain language for security reviews and first-time users.

What agents remember