Memory Curation

Memory curation is how agents organize raw episodes into structured knowledge. Unlike previous versions where an LLM did this automatically, Remind v1.0 puts the agent in control via snapshot and apply.

Agent-driven approach

The agent decides:

• What episodes to group into concepts
• How to resolve fact conflicts
• When to supersede outdated information
• What relations exist between concepts

Remind provides the tools; the agent provides the judgment.

The curation workflow

1. Capture experiences

During work, store raw experiences:

remember(content="Rate limiting is at gateway level", topic="architecture")
remember(content="Max connections is 100", episode_type="fact", entities="tool:postgres")

2. Review pending state

Use snapshot to see what needs curation:

snapshot(scopes="pending,conflicts")

Returns:

• pending — Unprocessed episodes with their entities
• conflicts — Open contradictions awaiting resolution

3. Curate with apply

Send a changeset that organizes the knowledge:

concept from=ep:11,ep:12 title="Gateway handles rate limits" "All rate limiting at gateway"
resolve id=conflict:3 winner=fact:abc "newer config value"
processed ids=ep:11,ep:12

Fact handling (automatic)

While concepts require agent curation, fact episodes are processed automatically:

1. Store — Episode is stored and embedded
2. Cluster — Assigned to existing fact cluster by entity overlap (Jaccard similarity ≥ 0.5), or a new cluster is created
3. Detect collisions — Active facts in the cluster with overlapping entities are flagged
4. Return result — Collision info is returned for agent triage

Collisions are NOT auto-resolved. The agent decides via apply:

# New fact supersedes old
supersede old=fact:abc new=fact:def

# Or: both are valid in different contexts
dismiss id=conflict:7 "staging vs prod"

# Or: genuine contradiction
conflict fact_a=abc fact_b=def severity=high

What curation produces

Given these episodes:

"User mentioned they like Python on Tuesday"
"User mentioned they like Rust on Thursday"  
"User mentioned they like TypeScript last week"
"User values type safety in all their projects"

The agent might create:

concept from=ep:1,ep:2,ep:3,ep:4 title="Language preferences" "User is a polyglot programmer drawn to statically typed languages"
link source=$c1 type=implies target=concept:compiled-languages
processed ids=ep:1,ep:2,ep:3,ep:4

The difference from automatic consolidation: the agent has full context of what these episodes mean in the current project, and can make informed decisions about how to organize them.

Pending vs processed

Episodes have a consolidated flag (semantically: "processed"):

• Pending — Not yet reviewed by the agent
• Processed — Agent has reviewed and either created a concept or marked as processed

Use snapshot(scopes="pending") to see what needs attention.

Use processed ids=ep:1,ep:2 to mark episodes as reviewed without creating a concept.

Fact clusters

Fact episodes are automatically grouped into fact_cluster concepts:

• Cluster title — Generated from shared entities
• Active facts — Facts with open validity windows
• Superseded facts — Facts that have been replaced (closed validity window, links to replacement)

Use snapshot(scopes="concept:cluster_id") to see full cluster detail including supersession history.

Time-travel

Fact clusters support time-travel queries:

recall(query="cache config", as_of="2024-06-01")

This shows what facts were valid at that point in time, not the current values. Useful for debugging "what did we believe then".