EchoDeskTrack: MemoryAgent2K memory budget

Chief-of-Staff memory that visibly remembers and forgets.

A persistent-memory agent demo for solo founders: memories are chunked, scored, embedded, retrieved by relevance + recency + importance + decay, then faded or consolidated over simulated time.

Simulated day

Qwen adapter

fallback

Chat interface

Cross-session Chief-of-Staff

Persistent memory across independent sessions.

Start by loading the demo facts.

Then ask from Session 2 to prove cross-session recall under the small context budget.

Memory Brain

Visible memory lifecycle

nodes brighten on recall
client
deadline
preference
decision
summary
small talk
fact

Active memories

0

Forgotten

0

Merged sources

0

Avg decay

0.00

Stored chunks

Decay curves and status

No memories yet. Load the demo or tell EchoDesk a preference, client detail, decision, or deadline.

Quantitative proof

Recall under limited context

precision

Avg context used

Run benchmark

Budget cap

2K tokens

The benchmark scores four fixed queries, including one that should not retrieve stale small talk after decay.

Lifecycle log

Encode → recall → decay → consolidate

Lifecycle events will appear here as the demo runs.

Architecture path

Chat UI → memory API service → Qwen-compatible scoring/embeddings adapter → PostgreSQL memory store → decay/consolidation job → Memory Brain.

Efficient retrieval

Top-K memories only, capped to 2000 estimated tokens.

Timely forgetting

Decay lowers inactive low-importance nodes and drops stale noise.

Sleep consolidation

Related memories merge into summaries with source lineage.