Open-source alpha research project

Introducing OOMP

Memory beyond the context window.

Persistent history. Selective recall. Finite prompts.

An independent, local-first, provider-neutral memory engine. Preserve source-backed history and selectively recall the evidence your AI agent needs.

Explore the Architecture
Illustrative memory flow Keep the history. Recall what matters.
BUILT AT INDUS AUTOMATION LABSLOCAL-FIRST PROVIDER-NEUTRAL SOURCE-BACKED

A past worth keeping.
A context worth selecting.

Follow a conversation beyond the active window.
From original source to the next finite prompt.

Skip visual story

01WORKING CONTEXT

A conversation moves on.
Its history shouldn’t disappear.

Every model prompt has a finite capacity. As a conversation grows, earlier details can fall outside the active context, even when an application keeps the transcript.

Longer windows and summaries can help. Durable memory adds another layer: retain the source, then select the relevant evidence when it is needed again.

Outside the active prompt does not mean deleted history.

The useful question isn’t “How much can we fit?”
It’s “What do we need to recall?”
THE LIVING MEMORYIllustrative demo — fictional data
AI HOST / ACTIVE CONTEXTFINITE WINDOW
  1. demo:evt-01

    I’m developing a project called Lumen.

  2. demo:evt-02

    I prefer concise technical explanations.

  3. demo:evt-03

    I plan to build a tabletop robot.

  4. demo:evt-04

    We discussed an enclosure for the planned tabletop robot. No decision was made.

  5. demo:evt-05

    Now, let’s work on a different task.

  6. demo:evt-06

    What should we explore next?

Earlier context passes this boundary

The prompt moves on. The source can remain.

A finite active window is not a claim that the host deletes its transcripts.

01 / Conversation / finite working space

02PERSISTENT HISTORY

Keep the source.
Build the memory.

Authorized source events are durably recorded first. Mind processing proposes memories; source and policy validation govern what is retained as attributed memory.

A message can remain part of the history without becoming a canonical fact. The original event remains the reference.

Source intake → durable event → Mind processing → attributed memory.

THE LIVING MEMORYIllustrative demo — fictional data
AUTHORIZED SOURCE INTAKEOriginal events recorded first
PROJECTProject Lumen↳ demo:evt-01
PREFERENCEConcise explanations↳ demo:evt-02
Some passages stay source-only.
  1. Source intake
  2. Durable event
  3. Mind + validation
  4. Attributed memory
Not every source passage becomes a canonical fact.

02 / Ingestion / source before interpretation

03THE OBJECT MAP

Objects,
with a past.

A project. A technical preference. A plan that hasn’t happened yet. A discussion with no final decision. Different kinds of history, connected without losing their source.

Explore the constellation. Hover or focus a node to preview its attribution; select it to keep the source open.

Associations organize evidence. They do not establish truth.

THE LIVING MEMORYIllustrative demo — fictional data
SOURCE INSPECTOR

Every memory has a way back.

Hover or focus a node to inspect its fictional source. Select to keep it open; Escape or Clear selection releases it.
Read the fictional source references
Project Lumen · demo:evt-01
“I’m developing a project called Lumen.” A project claim attributed to an original fictional user event. Attribution is not independent verification.
Concise explanations · demo:evt-02
“I prefer concise technical explanations.” An explicitly stated technical preference, with its own original source reference.
Tabletop robot · demo:evt-03
“I plan to build a tabletop robot.” A prospective activity: planned, not completed. The original wording preserves that distinction.
Enclosure discussion · demo:evt-04
“We discussed an enclosure for the planned tabletop robot. No decision was made.” An attributed discussion episode. The topic does not imply a finalized decision.

03 / Object relationships / inspect a source

04SELECTIVE RECALL

Recall
what matters.

An authorized question enters scoped retrieval. The project and its original source are selected in this example; unrelated objects stay in the archive.

OOMP returns a bounded evidence packet, not the final answer. Actual retrieval can miss information or return imperfect candidates.

The illustrated selection is predetermined, not a live retrieval result.

THE LIVING MEMORYIllustrative demo — fictional data
AI HOST / AUTHORIZED QUERY

“What project was I working on?”

SCOPED RETRIEVAL
Project Lumendemo:evt-01
SELECTED EVIDENCEProject Lumen↳ demo:evt-01source-backed · bounded
Still in the archive

Concise explanations · Tabletop robot · Enclosure discussion

FINITE EVIDENCE BUDGETSchematic capacity · not a token measurement

Selected in this illustration, not a live retrieval result.

04 / Retrieval / source-backed evidence selection

05BACK TO THE HOST

An expanding history.
A finite prompt.

Selected evidence crosses back to the AI host. Its language model combines that evidence with the current request inside a finite prompt.

Persistent history stays outside the model. Answer generation stays with the host. Neither storage nor model context is infinite.

Persistent storage ≠ the model’s native context window.

THE LIVING MEMORYIllustrative demo — fictional data
PERSISTENT HISTORY
OOMPOutside the model
SELECTED EVIDENCEProject Lumen↳ demo:evt-01source-backed · bounded
EXTERNAL AI HOST

Host model

FINITE MODEL PROMPT
Current request
Selected evidenceProject Lumendemo:evt-01
Room for the host’s context
Answer generation stays here.

An expanding history. A finite prompt.

Evidence is an input to the host, not an answer generated by OOMP.

05 / Handoff / the host model remains external

THE LIVING MEMORYIllustrative demo — fictional data
AI HOST / ACTIVE CONTEXTFINITE WINDOW
  1. demo:evt-01

    I’m developing a project called Lumen.

  2. demo:evt-02

    I prefer concise technical explanations.

  3. demo:evt-03

    I plan to build a tabletop robot.

  4. demo:evt-04

    We discussed an enclosure for the planned tabletop robot. No decision was made.

  5. demo:evt-05

    Now, let’s work on a different task.

  6. demo:evt-06

    What should we explore next?

Earlier context passes this boundary

The prompt moves on. The source can remain.

A finite active window is not a claim that the host deletes its transcripts.
AUTHORIZED SOURCE INTAKEOriginal events recorded first
PROJECTProject Lumen↳ demo:evt-01
PREFERENCEConcise explanations↳ demo:evt-02
Some passages stay source-only.
  1. Source intake
  2. Durable event
  3. Mind + validation
  4. Attributed memory
Not every source passage becomes a canonical fact.
SOURCE INSPECTOR

Every memory has a way back.

Hover or focus a node to inspect its fictional source. Select to keep it open; Escape or Clear selection releases it.
Read the fictional source references
Project Lumen · demo:evt-01
“I’m developing a project called Lumen.” A project claim attributed to an original fictional user event. Attribution is not independent verification.
Concise explanations · demo:evt-02
“I prefer concise technical explanations.” An explicitly stated technical preference, with its own original source reference.
Tabletop robot · demo:evt-03
“I plan to build a tabletop robot.” A prospective activity: planned, not completed. The original wording preserves that distinction.
Enclosure discussion · demo:evt-04
“We discussed an enclosure for the planned tabletop robot. No decision was made.” An attributed discussion episode. The topic does not imply a finalized decision.
AI HOST / AUTHORIZED QUERY

“What project was I working on?”

SELECTED EVIDENCEProject Lumen↳ demo:evt-01source-backed · bounded
FINITE EVIDENCE BUDGETSchematic capacity · not a token measurement
PERSISTENT HISTORY
OOMPOutside the model
SELECTED EVIDENCEProject Lumen↳ demo:evt-01source-backed · bounded
EXTERNAL AI HOST

Host model

FINITE MODEL PROMPT
Current request
Selected evidenceProject Lumendemo:evt-01
Room for the host’s context
Answer generation stays here.

An expanding history. A finite prompt.

Evidence is an input to the host, not an answer generated by OOMP.

Keep the history. Choose the evidence.

Continue to the implementation
Inspect the system boundaries

Memory has a place. Outside the model.

Ingestion and recall are separate paths. Source history stays in the durable store; selected evidence returns to the host. The trusted Python API also accesses the engine directly. MCP is optional. Retrieval permission does not grant ingestion authority.

SYSTEM OVERVIEWINGEST / RECALL

01 / INGEST

Preserve authorized history
  1. 01AI hostapproved source events
  2. 02MCP interfacesigned ingest authority
  3. 03OOMP engineMind extraction + validation
  4. 04Durable storelocal SQLite + source proofs

02 / RECALL

Return evidence to the host
  1. 01AI host → MCPauthorized recall request
  2. 02OOMP + storescoped search + selection
  3. 03Selected evidencebounded packet via MCP
  4. 04Host modelfinite prompt → answer
OOMP supplies evidence. The host’s language model generates the answer.
  1. Preserve the source

    Authorized events retain their original source references, roles and timestamps.

  2. Extract with attribution

    Rules, or explicitly configured local-model proposals, produce candidates checked against source spans and policy.

  3. Keep a durable history

    Local SQLite stores source events, attributed memories, episodes and temporal relationships.

  4. Recall within a budget

    Scoped retrieval selects relevant evidence and provenance for the host’s finite prompt.

Built around evidence.
Designed for control.

v0.1.0 / ALPHA

History that survives the session

Durable SQLite source events and attributed memories let relevant earlier work remain available after the recent conversation moves on.

SQLite · local persistence

A path back to the source

Event references and validated excerpts make returned evidence inspectable. Attribution explains where a claim came from; it does not establish truth.

Event IDs · source spans

Recall with boundaries

Lexical search, optional local embeddings and bounded graph traversal select evidence across sessions. Context packets have explicit byte budgets.

FTS5 · optional hybrid retrieval

Objects, relationships, time

Inspect a source-backed object map with multiple topic paths, typed relationships and temporal history. Building the projection is an explicit operation.

Object map · temporal graph

Your host. Your model.

An optional MCP interface connects authorized hosts through stdio or authenticated loopback HTTP. The host’s model stays outside OOMP and generates the answer.

MCP · separate read and write authority

Local operation, visible state

CLI onboarding previews profile setup. Owned-process supervision, diagnostics and inspection commands help operators understand their local system.

Onboard · doctor · lifecycle

Start local.
Inspect the source.

Install a reviewed source snapshot, then run a small fictional example. Core OOMP needs no TRACE, MongoDB, GPU or answer model.

Runtime
Python 3.11 / 3.12
Storage
SQLite with FTS5
Distribution
GitHub source
Read the complete setup guide
WINDOWS / POWERSHELL
01 — INSTALL + VERIFYPINNED / bebee61 ↗
git clone --branch main --single-branch https://github.com/hamzah105/OOMP.git
cd OOMP
git checkout bebee6194c0c5f3963b15cbbf56ebc10dd83ccf6
py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install ".[mcp,ui]"
.\.venv\Scripts\oomp.exe --version
.\.venv\Scripts\python.exe examples/quickstart.py

The checkout is pinned to the source reviewed for this page. The fictional quickstart ↗ prints an attributed memory and source IDs, reopens the store to verify persistence, then removes its temporary database. It makes no model calls.

POSIX shell equivalent
git clone --branch main --single-branch https://github.com/hamzah105/OOMP.git
cd OOMP
git checkout bebee6194c0c5f3963b15cbbf56ebc10dd83ccf6
python3.11 -m venv .venv
.venv/bin/python -m pip install ".[mcp,ui]"
.venv/bin/oomp --version
.venv/bin/python examples/quickstart.py

Shell equivalents are source-based; this website validation did not independently test Linux or macOS installation. See upstream platform coverage. Python 3.12 can be selected instead of 3.11.

Next: configure local MCP services

Optional: after the demo, onboarding previews profile configuration and requests approval before creating a database or credentials. Review the setup guide and choose a supported host.

.\.venv\Scripts\oomp.exe onboard --profile local
.\.venv\Scripts\oomp.exe doctor --full --profile local
.\.venv\Scripts\oomp.exe up --profile local

up runs in the foreground; Ctrl+C stops owned services. The last command starts local MCP and Mind processes. It does not host your answer model.

For an existing store, follow the read-only memory inspection guide ↗. It explains source events, derived memories and the separately built object map.

These commands reproduce bebee61 ↗. The latest main branch ↗ moves independently; review its documentation before switching.

No GitHub release published at the source review on 14 November 2026. Releases ↗ · Source history ↗

Source installation, not an advertised PyPI release or cloud installer. Read the current limitations ↗ before using personal data.

An expanding past. A finite prompt.

The long-term objective is memory that can navigate an increasingly extensive history without loading the entire archive into a model’s prompt. Better natural-language recall, evaluation and independent usability remain ongoing research.

Ongoing research · no literal infinite-context claim
Research goal

Broader natural-language recall

Evaluate acquisition across more phrasing, uncertainty and changing project state, with independently reviewed labels.

Read the research direction ↗
Research goal

Independent evaluation

Extend reproducible evidence studies with independent fixtures, exact tokenizer accounting and review of downstream answers.

Read the research direction ↗
Research goal

Measured operation at scale

Measure storage, latency and recovery across representative workloads. A growing archive still needs finite retrieval and prompt budgets.

Read the research direction ↗

No benchmark result is presented on this page. These goals describe future evaluation, beyond the implemented capabilities above.

Give your next session
something to build on.

Explore OOMP on GitHub
Understand the current boundaries
  • Alpha research software, without a production-readiness certification or universal host compatibility claim.
  • Local storage is plaintext SQLite. Protect the filesystem and backups; local-first does not mean encrypted.
  • Extraction and retrieval can miss relevant information or return imperfect candidates. Provenance is not factual verification.
  • Embedding and extraction models are optional and explicitly configured. Model capacity and retrieval work remain finite.
  • Evidence budgets count UTF-8 packet bytes, not exact model tokens. The host must also budget its own prompt and protocol overhead.
  • The documented submission profile uses one principal per namespace. Independently owned hosts need separate scopes and credentials.
Read the upstream limitations