ContextFit demo lab · memory retrieval without embedding APIs ★ GitHub
interactive memory demo

Nearest text is not always the right memory.

Agents do not just need search. They need to remember preferences, unresolved loops, changing facts, and scattered evidence — then explain exactly where the answer came from.

memory-machine.local
user asks later
“What should I remember before booking dinner?”
avoidheavymealsbeforeflightslikesquietpatioopen_loopFridayconstraintpreference
ContextFit retrieves

“Choose something light and low-chaos — they said travel days go better that way.”

Evidence is cited from the actual conversation, not hidden inside vector distance.

demo arcade

Four tiny memory traps.

Each one is a thing real agents get wrong when memory is treated as generic semantic search.

ASK

Nearest-text retrieval

plausible miss

ContextFit

cited memory
what the demo is showing

Memory has shape.

ContextFit keeps that shape visible: tokens, chunks, metadata, memory atoms, and evidence-ranked episodes.

01

Typed memories

Preferences, goals, constraints, decisions, open loops, and temporal updates are different signals.

02

Local token index

No embedding API is required. Retrieval runs on token-native structures you can inspect and move.

03

Structure-aware files

Markdown sections, TMD rows, JSON events, CSV rows, emails, and calendar events stay source-verifiable.

04

Evidence first

The result is not just a score. It is a ranked answer path with the chunks that justify it.

Build memory your agent can explain.

ContextFit is open source, local-first, and designed for the messy continuity layer real assistants need.