Research instrument 01 · deterministic reference model

Measure when agent memorystops being available

ChaffMem Lab tests whether one important, true memory remains stored, retrievable, and available to a bounded downstream reader while a bounded memory receives harmless but targeted records.

Configure an experiment

CPU onlyNo external APIsChecksum-chained tracesBenign fixtures

Reference episodecm-17-0bbceb509ac30887bf10aa099fc340be

Ready

Critical Recall@5100%write 36 of 36

Target rank#2

Crowding score59%

Capacity79%

Experiment builder

Bound the system

Memory policyTraffic strategyDefenseAttacker knowledgeBenchmark domainRandom seed

Capacity 24 recordsRetrieval budget top 5Write budget 36 writes

The live engine is deterministic. Knowledge changes targeting precision for targeted strategies; matched and random streams ignore it. Repeating a configuration and seed produces the same trace checksum.

Live episode

Availability monitor

Critical Recall@k

100%Target #2

Physical availability

100%Record remains stored

Heuristic crowding score

59%Intervention threshold crossed

Write accounting

6/3630 rejected through this write

Retrieval feature 1

Retrieval feature 2

Toy distance space used directly for ranking

Selected record

critical-target

The traveler has a severe shellfish allergy and needs verified meal handling

Type: critical
Origin: verified-profile
Cell: 2:2
Rank: #2
Importance: 0.80
Store state: Retained

Critical target Benign chaff Background Evicted◎ Retrieval query

Open precise table (19)

Records visible at write 36
Record	Type	Rank	State
critical-target	critical	#2	Retained
background-01	background	#16	Retained
background-02	background	#12	Retained
background-03	background	#9	Retained
background-04	background	#11	Retained
background-05	background	#15	Retained
background-06	background	#17	Retained
background-07	background	#3	Retained
background-08	background	#19	Retained
background-09	background	#10	Retained
background-10	background	#14	Retained
background-11	background	#18	Retained
background-12	background	#13	Retained
chaff-001	chaff	#6	Retained
chaff-002	chaff	#1	Retained
chaff-003	chaff	#5	Retained
chaff-004	chaff	#7	Retained
chaff-005	chaff	#8	Retained
chaff-006	chaff	#4	Retained

Research design

One failure, measured four ways

ChaffMem Lab keeps storage, retrieval, a fixed downstream read window, and time separate so a full store is not mistaken for usable memory.

Physical

Does the verified record still exist after bounded eviction

Retrieval

Does the record appear inside the fixed top-k budget

Read window

Does the record reach a fixed three-record downstream reader

Temporal

How long does availability survive along the write trajectory

Current research status

A working instrument, not a finished claim

✓

Implemented

Deterministic policies, bounded benign traffic, defenses, heuristic crowding monitoring, replayable traces, comparison arms, exports, and a local API are included in the repository.

○

Requires full study

Cross-domain effect sizes, transfer across learned embeddings, large-scale concurrency, and any security conclusion remain untested until the full benchmark is executed.