Diop Daily #137 — October 2026

A Memory Must Be Able to Leave

An agent can remember a preference, a procedure, a correction, or the history of a decision. That memory becomes institutional capital only when the institution can move it. A runtime that stores months of learned context but cannot export, inspect, revoke, or re-hydrate it elsewhere has created continuity for a tenant, not durable capability for an owner.

The distinction is becoming visible in public technical work. The W3C AI Agent Memory Interoperability Community Group adopted a charter describing a vendor-neutral shape for agents to store, recall, share, and erase memory across implementations. Its stated scope includes portable memory cells, post-quantum identity, capability-scoped sharing with revocation, tamper-evident audit trails, and cryptographic erasure. The group is a Community Group incubator, not a W3C standard, and its charter does not certify an implementation. It does, however, name the missing layer with useful precision: memory needs a transfer contract.

A memory that cannot leave the system that produced it is a dependency disguised as intelligence.

Continuity belongs to the institution

Agent memory is often discussed as a retrieval problem. Teams ask how many facts can be recalled, how quickly a vector store can answer, or whether a long context window improves coherence. Those questions matter inside one runtime. They leave ownership unresolved.

Institutional memory has a wider life. It includes the reason a decision was made, the source that supported it, the person or office that supplied it, the procedure that followed, the correction that superseded it, and the permissions that limit its use. It may contain several forms at once:

  • Episodic memory: what happened, when it happened, who participated, and what evidence was present.
  • Semantic memory: a claim, definition, relationship, or local category that an agent may use later.
  • Procedural memory: a tested way of carrying out work, including prerequisites, refusal conditions, and recovery steps.
  • Working memory: the temporary state of a task, its open questions, and the boundaries that should expire with the assignment.
  • Identity memory: the principals, mandates, roles, and institutional relationships that make a piece of context trustworthy in a particular setting.

These categories cannot all be transferred as plain text. A sentence extracted from a prior conversation may look useful while losing its date, author, jurisdiction, confidence, or status. A procedure without its refusal conditions can turn a lesson into an unsafe command. A preference without a principal can become an unowned instruction. Portability therefore means preserving the relationships around memory, not merely copying the words.

Export is a constitutional act

When a vendor offers an export button, the visible file is only the beginning. A serious export must answer five questions: what was selected, who authorized the selection, how was its integrity protected, where did each item come from, and how may the receiving system use it?

A June message on the W3C memory-interoperability list provides a concrete reference point. It discusses a proposed implementation with ML-DSA-65 identity binding, Merkle provenance, partial-disclosure proofs, per-cell encryption, cryptographic erasure aligned with GDPR Article 17, and capability-scoped sharing. The message is an open technical discussion, not a standards approval. Its value lies in making the design pressure inspectable. Selective transfer needs cryptographic evidence and policy boundaries because a memory store is rarely safe to disclose as one undifferentiated block.

The portability problem also appears in research. The May 2026 preprint Portable Agent Memory proposes a five-part model spanning episodic, semantic, procedural, working, and identity memory. It combines content-addressed provenance, capability-based access, and a re-hydration pipeline intended to reduce memory-mediated prompt injection. Its reported pilot results are author claims from a preprint rather than a field-wide benchmark, while the architecture remains useful because it treats transfer as a protocol boundary with its own threat model.

Export rules should make the following visible:

  • Selection: which memory cells crossed the boundary and which remained behind.
  • Lineage: the events, sources, transformations, and revisions that produced each cell.
  • Authority: the principal, institution, or delegation that permitted transfer and the purposes covered by that permission.
  • Integrity: hashes, signatures, versions, timestamps, and proofs that let the receiver detect alteration or omission.
  • Erasure: how a cell can be revoked or deleted, including the consequences for derived summaries and cached copies.
  • Re-hydration: how the receiver frames memory as evidence rather than executable instruction, and how it records what it accepted.

This is close to constitutional design. The export defines what an institution is willing to carry forward, what authority accompanies it, and what the next runtime is forbidden to infer. It is a boundary document between past experience and future action.

Interoperability without memory is shallow cooperation

Google’s Agent2Agent announcement describes agents collaborating across vendors and frameworks, with support for capability discovery, task management, messages, artifacts, and long-running state updates. Its design principles explicitly allow agents to collaborate even when they do not share memory, tools, or context. That separation is sensible: a remote agent should not receive an entire private memory store merely because it has been asked to perform a task.

It also exposes the limit of task interoperability. A message can delegate an assignment, but the receiving agent still needs the relevant history, definitions, constraints, and identity context. If that context is reconstructed informally each time, the organization pays for translation at every boundary. If it is transferred wholesale, privacy and injection risk grow. The missing design is a scoped memory handoff: enough inherited context to make the task intelligible, with evidence and permissions that can be checked independently.

For African institutions, this boundary has political weight. A ministry, university, archive, or cooperative should be able to change a model provider without exporting its accumulated institutional knowledge as an opaque vendor-specific database. It should be able to share a local-language procedure with a regional partner while keeping unrelated personnel or community records private. It should be able to prove why a translated term entered a workflow and who can correct it. A memory format that preserves only the dominant platform’s categories quietly turns sovereignty into subscription dependence.

The investable object is the transfer contract

The commercial object is a portable memory-transfer layer: schemas, registries, cryptographic evidence, policy enforcement, migration tooling, and re-hydration tests that let institutions carry selected capability across runtimes without carrying hidden risk.

  • Memory registries: stores that preserve typed cells, provenance, language, jurisdiction, retention, and custodianship rather than presenting memory as an undifferentiated vector index.
  • Export and migration tooling: adapters that map vendor-specific memories into a documented interchange format, report loss or ambiguity, and produce a migration manifest.
  • Policy-aware disclosure: systems that issue scoped, expiring access to selected memories and record revocation, downstream copies, and derived artifacts.
  • Integrity and erasure services: signing, hash proofs, audit trails, deletion propagation, and independent checks that make the transfer legible to a regulator or counterparty.
  • Re-hydration evaluation: tests for fidelity, authority recognition, language preservation, prompt-injection resistance, refusal quality, and safe degradation when a memory cell cannot be trusted.

Capital should examine measures that expose ownership rather than dashboard activity: the percentage of memories with complete lineage; the share that can be exported without silent loss; the time required to revoke a cell across derived stores; transfer fidelity by language and memory type; successful reproduction of a decision from the exported record; and the number of provider migrations completed without losing policy, provenance, or operating knowledge. These are measures of institutional continuity.

The strongest companies in this layer will show exactly what can move, what cannot, why the boundary exists, and how the receiving system behaves when evidence is incomplete. That discipline is commercially important. A migration tool that hides loss creates a new form of lock-in under the language of portability.

Build for departure before you need it

An institution can begin with one consequential memory class: approved procedures, research annotations, customer preferences, or public-service case history. Define the owner and the retention rule. Give every cell a type, source, timestamp, language, authority, and confidence. Separate observations from instructions. Sign the export. Test selective disclosure. Delete a cell and verify what disappears downstream. Then re-hydrate the surviving record in a second runtime and ask an independent reviewer whether the resulting agent knows what it is allowed to know.

That final question is the heart of the matter. Memory should increase continuity without dissolving judgment. A receiving agent needs inherited knowledge, but it also needs the right to doubt, ask for missing evidence, refuse an expired instruction, and identify a conflict between an old procedure and a current authority. Portability earns trust when it carries both capability and the conditions under which capability may be used.

African intellectual sovereignty has always required more than possessing archives. It requires the power to preserve, interpret, transmit, and correct knowledge under institutions that can answer for their choices. Agent memory will follow the same law. If it remains trapped inside the vendor that accumulated it, it strengthens dependency. If it can leave with its lineage, permissions, languages, and correction path intact, it becomes a portable instrument of institutional freedom.

Sources