PureTens⊗rLab
Part of pureWRIGHT · Engineering Record

pureMIND

The part that remembers everything.

Every brief, note, decision, and lesson you give the system stays with it. Ask next month and it picks up where you left off. pureMIND is the one record behind that promise: every part of the system reads from it and writes back to it, each entry with its source and its trust attached. It runs today underneath PureTensor's own operations, and its canonical store is plain files in a repository you can read, move, or delete. Below is how it is built, so you can judge the claim rather than take it on trust.

STATUS: OPERATIONALDOCUMENT REV 2026-10-06FIGURES AS OF 2026-10
The Film · 3 min 22 s

How our agents remember

Capture, one writer, two indexes, a three-way recall, memory that keeps its own history, and the overnight work that keeps it honest. Every frame was generated by code and rendered on our own GPUs; the figures on screen are measured, and the names are illustrative.

Benchmark
On LongMemEval_s, a public 500-question long-term-memory benchmark, an agent reading pureMIND's retrieved context answered 90.6% correctly with 25 retrieved passages (median 11.7k tokens of context) and 87.8% with 10 (median 5.0k tokens), at 98 ms p95 retrieval latency (sequential queries) on local hardware. Reader and judge: OpenAI gpt-5.4. Same retrieval pipeline design as production (hybrid, fused, re-ranked), measured in August 2026.
Boundary
Agents send the context they retrieve to the model they are running on. A disaster-recovery index of notes and session reports is held with a cloud search provider, with no scheduled refresh. Several maintenance jobs use a local model first and a hosted model as fallback, and two overnight jobs (the nightly reflection and an operations job) run on a hosted model.
Models
Embedding and re-ranking use open-weight models, run on hardware PureTensor owns.
Section 01 · The Problem of Forgetting

Every session used to start with a stranger

A language model wakes with an empty context window. Whatever it learned yesterday, whatever broke last month and how it was fixed, whichever names, decisions, and dead ends the operator has already paid for: all of it is gone at the next session boundary.

For a chatbot that is an inconvenience. For autonomous infrastructure it is disqualifying. An operations agent that cannot remember re-derives its own runbooks, repeats its own outages, and asks its operator the same question every week. Continuity is not a feature of an autonomous system; it is the precondition for one.

pureMIND is the answer PureTensor runs in production: one memory substrate that captures what every session, incident, and voice note taught, consolidates it into durable topics, and returns the relevant slice to whichever agent needs it next.

1,308
Durable topics
1,561
Graph entities
7
Memory tools
2,982
Live relationships
Section 02 · The Write Path

How an observation becomes knowledge

What the fleet learns lands as plain markdown under version control through one writer. Agents edit notes or propose changes to a ledger, the writer commits both every 30 seconds, a nightly job promotes the day's log into always-loaded memory, and the commit history keeps every version. Session transcripts, reports and voice notes are indexed for search only.

Who writes
  • Agent sessions
  • Operator corrections
  • Nightly jobs
  • Other agents
Stage 01Single writerledger proposalsdirect editspropose-write gateCommits to git every 30 seconds.
Plain filesGit history is the audit log.
  1. Stage 02Canonical topicsgit-versioned markdownReviewed nightly: duplicates and broken links flagged for the operator.
  2. Stage 03Ranked mirrorrecency × centralitysize-boundedAn index of titles, loaded at session start.
OutputSession context
Section 03 · Recall

Three lanes, one answer

Every query runs three lanes across the memory corpus, keyword, semantic and summary, fuses the candidates, and reranks them on GPU. Results are then weighted by source trust and recency. When a model server is unavailable, retrieval steps down a ladder instead of failing.

InputQuery
Vault notes, summaries and reports
Vault notes, summaries and reports
  • Keyword lanePostgres full-textno GPU needed
  • Semantic laneembeddingsGPU-servedpgvector
  • Summary lanenightly summariesvector index
  1. FusionReciprocal-rank merge of all three lanes.
  2. Rerankcross-encoderGPU-servedtop 50
  3. Weightingrelevancesource trustrecency
OutputInjected context
Failover ladderDegrades when a model server fails.
  1. GPU
  2. CPU standby, same model
  3. Keyword only
Section 04 · The Immunity Loop

Why a failure handled once can be recalled

pureMIND closes the loop. When an agent resolves an incident it writes the lesson down as a note; the note is indexed, and the next agent that investigates the same shape finds it before it starts.

pureMINDThe notes close the loop
  1. 01IncidentAn agent investigates
  2. 02DiagnosisRoot cause found
  3. 03FixApplied and checked
  4. 04LessonWritten as a note
  5. 05IndexedSearchable across the corpus
  6. 06RecalledWhen an agent next investigates
A failure handled once becomes a lesson. The corpus only grows.
Section 05 · Sovereign by Construction

Memory you could walk away with

Sovereignty here is not a policy statement. It is a set of construction choices that make the memory portable, inspectable, and independent of any vendor, including us.

Plain files, under git

The canonical store is markdown in a version-controlled repository. No proprietary database holds the memory hostage; any tool that can read a file can read the corpus, and the commit history is the audit trail.

Inference on owned silicon

Embedding and reranking run on GPUs PureTensor owns. Embedding has a second on-premises endpoint and reranking a CPU fallback. If every embedding endpoint is unavailable, recall falls back to keyword search rather than failing.

One writer, full history

Agents propose changes to a ledger; one writer screens proposals from outside the trust boundary, then applies and commits every change, so every committed version of every note is kept. Contradictions and note merges are proposed as reviewable events; passages are retired from the search index automatically only when they are near-duplicates or their source note has changed.

Own boundary, named exceptions

Capture, storage and recall run on hardware PureTensor owns. What leaves that boundary is named: the context an agent retrieves goes to the model it is running on; several maintenance jobs use a local model first and a hosted model as fallback, and two overnight jobs (the nightly reflection and an operations job) run on a hosted model; and an off-site disaster-recovery index of notes and reports exists, with no scheduled refresh.

From the Lab

The research behind pureMIND

End of Document

Built, running, and yours.

pureMIND is the part of pureWRIGHT that remembers everything, alongside pureBOT, which does the work, and pureMEND, which repairs what breaks. It keeps what the system learns and what you tell it, in files that belong to you, and it is deployed for enterprise as the canonical institutional record. The engineering companion piece is the pureBOT record.

pureBOT recordTalk to our engineers