In My Pocket Labs

PROJECT 5.5 Full argument and decision record · August 2026

Memory that can evolve.
Work that remains accountable.

Research is teaching agents to navigate memory rather than wait for a search system to hand them context. We think that is the right direction for knowledge—and an incomplete operating model for a business.

Our position: let agents navigate evidence flexibly. Keep authority, ownership, approval, and verification explicit.

01

Our position

The hard problem is no longer simply “Can the system remember?” It is “Can the system use memory without quietly turning recollection, inference, or a message from the outside world into permission to act?”

THE RESEARCH HORIZON

Memory for questions we cannot predict.

Systems such as MemGPT, MemOS, and NapMem treat memory as an active resource. Instead of loading one large history or accepting a fixed retrieval result, an agent can move between summaries, structured records, and original evidence.

That direction matters because future agents will work across longer time spans, more tools, and changing goals.

THE OPERATING HORIZON

Controls for consequences we can predict.

A business already knows many of its dangerous failure modes: the wrong account, the wrong project, an unapproved message, a duplicated external action, stale source data, or an unverified result.

Those failures should not be left to semantic confidence. They need explicit rules, ownership, and evidence.

Active, provenance-linked memory for understanding. Deterministic, bounded, independently verified execution for action.

02

The system

One architecture, four distinct responsibilities.

The point is not to make everything deterministic. The point is to put flexibility where it helps and constraint where mistakes carry consequences.

01KNOWLEDGE

Find the right evidence

Conversations, typed records, topic tracks, project briefs, and source references connected by provenance.

02GOVERNANCE

Decide what is allowed

Business boundaries, project ownership, policies, procedures, approvals, and one-writer rules.

03EXECUTION

Do bounded work

Narrow owner agents act through approved tools against the systems that actually hold the facts.

04PRESENTATION

Make work legible

A human-facing Desk shows attention, decisions, progress, evidence, results, and verification.

The rule between the planes

Understanding a fact does not grant permission to act. Displaying a result does not make the display authoritative. A semantic match may suggest a route; the registered system must validate it.

03

Decision chain

We did not arrive here from a whiteboard.

This position grew through nearly a year of building, operating, failing, and correcting. The public version below removes private operating detail; the full evidence-backed record belongs in Control Center under Attempt 5.5.

  1. 01
    ATTEMPT 5.5

    A procedure has to survive the conversation.

    An early AI receptionist experiment exposed the difference between remembering instructions and reliably carrying a workflow through states, handoffs, exceptions, and verification.

  2. 02
    CONTROL CENTER

    Shared context needs an authority boundary.

    Central project history improved continuity, but also revealed that a project registry cannot replace the databases, repositories, and external services that hold operational truth.

  3. 03
    GLOBAL

    Some rules belong above every project.

    Privacy, business separation, and approval rules needed a portfolio-wide policy layer without turning that layer into another operational database.

  4. 04
    MCP

    Fast access is useful; unrestricted access is not.

    A typed doorway made orientation faster. Narrow, append-only writes proved safer than a general tool that could change anything it could reach.

  5. 05
    DANIEL'S DESK

    The interface should manage attention—not become a second brain.

    When the presentation layer began accumulating routing rules, duplicated state made the system harder to trust. The Desk was recast as an approval, progress, results, and verification surface.

  6. 06
    EMAIL + CLOUDFLARE

    An incoming message is an event, not an instruction.

    Mailbox polling was fragile, so direct inbound delivery became attractive. The design was refined: exact recipient routing, immutable receipt, duplicate checks, a human-readable archive, and no automatic authority for raw email or attachments.

  7. 07
    NAPMEM

    Flexible navigation belongs in the knowledge path.

    The paper supplied a clearer model for linked memory levels and active retrieval. It also sharpened our conclusion: learned navigation can improve what an agent knows without deciding what the agent is allowed to do.

04

The claim

Are we actually on the cusp?

Yes—as applied work. No—if “on the cusp” is taken to mean we invented the underlying ideas.

WHAT THE EVIDENCE SUPPORTS

The pieces are converging.

Research is moving from passive retrieval toward structured, navigable memory. Production guidance is moving toward scoped tools, guardrails, human control, and audit trails. Bringing those ideas together over real, multi-project business operations is still early.

WHAT WE SHOULD NOT CLAIM

This is not a solved architecture.

We have not proven a new general memory algorithm, eliminated model error, or demonstrated safe full autonomy. Our contribution is an operating thesis, a decision trail, and a system that can be evaluated against real failures.

WHAT WOULD CHANGE OUR MIND

A thesis should be able to lose.

  • If deterministic routing creates more error or delay than it prevents.
  • If active navigation cannot improve evidence selection under a bounded cost.
  • If provenance cannot survive summarization and memory promotion.
  • If human approval becomes ceremonial rather than meaningful control.
  • If a simpler architecture produces the same reliability with less operational burden.
05

Evidence

The work we are building from.

These sources do not prove our complete architecture. They establish the research direction and the production constraints we are trying to reconcile.

2026 · RESEARCHFrom Passive Retrieval to Active Memory Navigation

Linked memory levels, provenance, and a learned policy for choosing what to inspect.

Read paper ↗ 2025 · RESEARCHMemOS: An Operating System for Memory-Augmented Generation

Memory as a first-class, traceable, evolvable resource for future intelligent systems.

Read paper ↗ 2023 · RESEARCHMemGPT: Towards LLMs as Operating Systems

Hierarchical memory management and control flow across longer-running interactions.

Read paper ↗ PRODUCTION GUIDANCEA practical guide to building agents

Scoped tools, layered guardrails, human intervention, and oversight for high-risk actions.

Read guide ↗ PRODUCTION RESEARCHTrustworthy agents in practice

Human control, transparency, privacy, and permission boundaries as agent autonomy grows.

Read article ↗

OPEN FOR CRITIQUE This is a working position

Tell us where the argument breaks.

We are especially interested in counterexamples, prior art, failure modes, and simpler architectures that preserve both adaptability and accountability.

Send feedback