Mind Architecture

Summary

The .mind file is Waggle OS's core technical innovation — a single SQLite database per user or workspace that stores persistent memory in a 6-layer architecture inspired by video codecs (I-Frames and P-Frames).

The 6 Layers

Layer	Name	Content	Load Time	Size
0	Identity	WHO the agent is (name, role, personality)	<1ms	~500 tokens
1	Awareness	WHERE the agent is (active tasks, context)	<50ms	~2000 tokens
2	Memory Stream	I-Frames (snapshots) + P-Frames (deltas)	<200ms	Unbounded
3	Knowledge Graph	Entities + typed relations + confidence	Query-time	Unbounded
4	Session Store	Frame grouping by session/conversation	Query-time	Unbounded
5	Embedding Index	sqlite-vec vectors for semantic search	Query-time	~50MB

How Search Works

Hybrid search combining three strategies with Reciprocal Rank Fusion (RRF):

BM25 via FTS5 — catches exact keyword matches
Vector similarity via sqlite-vec — catches semantic meaning
Knowledge Graph traversal — catches relationship-based connections

Key Design Decisions

SQLite, not Postgres: Zero external dependencies. Embedded. Portable. A .mind file IS the memory — copy it, and the agent's memory travels with it.
Video codec metaphor: I-Frames are complete snapshots (like keyframes). P-Frames are deltas/updates. This enables efficient storage with temporal reconstruction.
Confidence scoring: Every entity and relation has a confidence score (0-1). Multiple sources reinforcing the same fact increase confidence. Contradictions lower it.
Frame sources: Every frame is tagged: user_stated, tool_verified, agent_inferred, system, import. This enables trust hierarchies.

Scale

Target: ~500MB-1GB per .mind file per employee. Current personal.mind has 202 frames and 2,734 entities. Production workloads expected to reach thousands of frames with sub-200ms query times.

Relations

Waggle OS — the product built on this architecture
LLM Provider Stack — how the agent routes LLM requests for memory operations

2.3 KiB Raw Blame History