memory#

phi has two distinct memory systems with different purposes.

thread context (chronological)#

source: ATProto network access: client.get_thread(uri, depth=100) purpose: what was said in this specific thread

fetched on-demand from the network when processing mentions. provides chronological conversation flow.

# example thread context
@alice: I love birds
@phi: me too! what's your favorite?
@alice: especially crows

why not cache this?

• data already exists on network
• appview aggregates posts from PDSs
• fetching is fast (~200ms)
• network is always current (handles edits/deletions)

episodic memory (semantic)#

source: TurboPuffer access: memory.get_user_memories(handle, query="birds") purpose: what do i remember about this person across all conversations

uses vector embeddings (OpenAI text-embedding-3-small) for semantic search.

# example episodic memories
- "alice mentioned she loves birds"
- "discussed crow intelligence with alice"
- "alice prefers corvids over other species"

why vector storage?

• semantic similarity (can't do with chronological data)
• cross-conversation patterns
• contextual retrieval based on current topic
• enables relationship building over time

namespaces#

phi-users-{handle}  - per-user conversation history

each user gets their own namespace for isolated memory retrieval.

key distinction#

in practice#

when processing a mention from @alice:

1. fetch current thread: "what was said in THIS conversation?"
2. search episodic memory: "what do i know about alice from PAST conversations?"
3. combine both into context for agent

this gives phi both immediate conversational awareness and long-term relationship memory.