Intent-relevant memory: a finite window over a growing archive

Status: scoped investigation and v0.2.2 implementation, September 18, 2026. This is not a claim of literal infinite model context, perfect recall, or a new Terminal-Bench score. No VM, task sweep or generation-model benchmark was run.

The incident and immediate fix

A real session stopped after 55 tool results, then stopped again on Resume with ContextOverflow: Context accounting reserve exhausted. In v0.2.1, context was trimmed to its budget before adding archived_execution_count. That extra JSON field could overflow an otherwise valid projection. Retrying rebuilt the same oversized state, rather than repairing it.

The fix counts every field before fitting. If necessary, optional retrieved pages, historical rows, duplicate source snapshots and older detailed effects are removed from the projection. Original records remain saved. Current intent, earlier user requests, previous conversational reference, latest effect and verification/status metadata are not silently discarded.

A read-only diagnostic projected the actual saved history through both versions while varying reserved-overhead bytes to locate the exact bookkeeping boundary:

Measurement	Result
Reserved overhead at matching boundary	1,973 bytes
Available router budget	22,027 bytes
Old projection	Exact reported reserve-exhausted exception
Corrected projection	21,022 bytes, 51 archived effects
Saved log changed / effects replayed	No / no

The overhead above is a controlled matching boundary, not a claim to have reconstructed the original process's exact MCP catalog. The user's trace and paths remain private and are not in this repository. No provider was called by this diagnostic.

What the literature supports

• MemGPT (2023): external storage and explicit paging can provide the appearance of more context while model calls remain finite. Its experiments do not establish perfect retrieval or unlimited task performance. This motivates separate authoritative archives and disposable working sets, not continually increasing the prompt.
• Lost in the Middle (2023): relevant evidence can be used less effectively when buried in long input. More supplied text is not necessarily better. Its tested models/tasks are not current JMP scores.
• LongMemEval (ICLR 2025): evaluates extraction, cross-session reasoning, temporal reasoning, updates and abstention, and separates indexing, retrieval and reading. This motivates testing newer corrections and missing evidence rather than just easy keyword needles. We did not run the LongMemEval benchmark or reproduce its published results.

Scoped Jev retrieval study

Source: experiment.py. The fixture contains one user record, 12 hand-authored fact records and 3,000 unrelated later tool records. There are 10 answerable questions and two intentionally unsupported questions. It covers exact terms, paraphrases, a prohibition, approval, an updated configuration and a resolved diagnostic. Original negative and superseded facts stay in the corpus.

Method	Measurement	Observed
Latest 8 records	Contains answer on 10 answerable questions	0/10
SQLite FTS5/BM25, top 4	Contains answer on 10 answerable questions	9/10
Candidate union: lexical top 16 + recent 8 + first 8 fact records	Contains answer	10/10
Jev-1.13.0, single choice over that candidate union	Correct record or correct abstention	12/12

The lexical miss was the paraphrase about whether an autonomous job could fetch packages from the internet; its fact concerned public-network access in unattended runs. Jev selected that fact when the diversity tier exposed it. Jev also chose newer Meridian/Falcon records and abstained on both unsupported questions.

Measured one run: 12 Jev calls, 31,791 reported input tokens, 2,044 reported output tokens, zero request errors; 1.957 seconds cumulative request time, 0.153 seconds median and 0.288 seconds maximum. These are sequential API timings on one machine/connection, not end-to-end agent latency or causal speedup. Provider billing cost was not measured; Jev is billable. No DeepSeek/Bonsai call was made.

Important caveats

This is a tiny, hand-authored, intentionally old-evidence fixture, not an independent quality benchmark. The early-history diversity tier favors this layout and cannot scale to every old fact. Recency's failure is expected by construction. Retrieval recall@4 and single-choice selection accuracy are different metrics: their percentages are not a fair head-to-head answer-quality comparison. There is no blind reader, randomized placement, embedding baseline, repeated sampling, adversarial instruction suite or real task-success measurement. The study does not establish calibrated confidence, general reliability, or a causal Jev contribution over other semantic rerankers. If the correct record is not in the candidates, Jev cannot select it.

What ships, versus what remains research

Shipping: jmp/memory.py builds a session-local SQLite FTS5 cache of original user messages, assistant finals and tool results. Large saved output is indexed in ~2 KB text pages with 128-character overlap, rather than only its head/tail excerpt. Event sequence, kind, status and page preserve provenance. Only artifact paths resolving inside that session are read. Search uses literal terms, not user-supplied SQL/FTS syntax. It never searches arbitrary workspace files.

Context can include up to two relevant older execution pages beside recent results, only if budget permits. Jev selects the recall tool when more historical evidence is needed; the generation model supplies a search query or exact record_seq, and next_offset supports further pages. This uses the same native-function validation boundary as other actions, but archive lookup itself is local and read-only. Recalled output is excluded from indexing to prevent recursive self-reinforcement. A recall does not invalidate current source caches or count as a possible workspace change. Missing search hits never imply absence.

Not shipping as a default: the study's extra Jev reranking call and fixed early-history candidate tier. One favorable synthetic run does not justify extra calls on every context build. Production uses local lexical selection and Jev's existing tool-routing decision to request more evidence; these are distinct from an always-on semantic reranker. No hidden extra memory-provider calls occur.

The cache memory.sqlite3 is disposable private data alongside the authoritative session log/artifacts. If damaged, optional automatic retrieval falls back to the existing context; explicit recall reports an error. Close the app before removing a damaged cache to rebuild it next time. Indexing/retrieval still consumes disk, CPU and RAM, and sessions themselves currently load their event log into memory. This is not constant-memory processing of an infinitely long log.

Hard limits we deliberately keep

All user requirements remain pinned, rather than guessed away by a relevance model. A task whose required instructions/schema/metadata alone exceed the budget can still pause safely. Arbitrarily long conversations with unbounded accumulating user instructions are therefore not solved by this release. Archives persist; model windows, disk and retrieval recall are finite. Old file observations may be stale, assistant claims are not evidence, and execution uncertainty is never cleared by retrieval. Actual current source still needs reading before edits.

The next defensible experiment is randomized evidence placement, distractor and paraphrase variation, withheld tasks, multiple relevant records, update/conflict cases and reader/task outcomes under equal token/cost budgets. Durable constraint revision needs its own design before dropping old user instructions automatically.

Reproduce only this investigation

# Local indexing and retrieval only; no provider calls
python research/context_memory/experiment.py
# Explicitly paid Jev-only 12-call selection experiment
python research/context_memory/experiment.py --live

Uses installed project dependencies. Credentials follow the application's private Jev vault/environment setup. Results are written under ignored .local/research/context-memory/; do not publish raw local traces. The script never executes a workspace task or reads the user's conversations/workspace code. A source build/static audit and this scoped study do not certify the new release on every Windows system; no broad test-suite run was performed for this update.