Append-Only with Retraction in Agentic Memory

Kendall Clark · Pentad Labs · 2 June 2026 · PLRN-007

Abstract

WunderOS began with Wunderblock, an agentic memory subsysten, and has grown in scope as the lack of inviolable adjacency markers in the emerging commercial space obviously rationalizes. The first and key insight: agenticy memory is a contextualized, multi-agent retrieval, not a storage, problem. However, agentic memory should never destructively delete. Staleness is a retrieval-time authority problem, solved by decay-weighted scoring and a currency signal carried with each retrieved fact, not a storage problem solved by erasure.

A published critique of append-only agent memory holds that lifecycle operations on memory (redaction, retirement, supersession, relaxation) are non-negotiable, because every stale artifact competes for attention, search rank, review time, or behavioral authority. The critique is correct about the failure and wrong about its remedy. Each of the four lifecycle operations is realizable without destructive deletion, and the attention-competition failure it names is real but lives at retrieval time rather than in storage. The exception is legal erasure, which logical retraction does not satisfy and which requires a destructive escape hatch that breaks the append-only invariant purposively and audits the break.

1. The critique, on its own terms

zby, in Designing Agent Memory Systems, argues that append-only memory is insufficient for agents, and that any serious memory system must support lifecycle operations: it must be able to redact a fact, retire a fact, supersede a fact, and relax a fact’s authority. The reason given is attention competition. Every stale artifact in the store competes for attention, search rank, review time, or behavioral authority, so a system that can only add records will accumulate stale records that degrade retrieval and mislead the consuming agent. On this view, append-only is a storage convenience purchased at the cost of correctness, and the lifecycle operations are the price of keeping the store usable.

We agree that the failure is real. A store that grows without bound and ranks every record equally will surface stale facts alongside current ones and let the stale ones win on recency of insertion or on lexical match. Yet the conclusion does not follow necessarily. The four operations are operations on what a query sees and what authority a hit carries, not operations on what the storage layer retains. They are realizable at retrieval time, and realizing them there preserves properties that destructive deletion forecloses.

2. Why append-only is the right default

WunderOS substrate commits to append-only memory with retraction tombstones and no destructive deletes. The commitment is not aesthetic. Three properties, each key to platform operations in regulated enterprises, depend on it. We deeply bias in favor of monotone operators as a result.

  1. Audit. A store that never deletes can answer what was known and when it was known, for any past instant, because the record of every assertion and every retraction is still present. Destructive deletion erases the answer along with the fact.
  2. Multi-tenant isolation. Deletion is a write that mutates shared structure, and mutation under concurrency across tenants is where isolation guarantees are hardest to keep. Append-only confines each tenant’s writes to additions, so the isolation argument reduces to controlling what each tenant may append and what each tenant may read.
  3. Deterministic replay. Replaying the event log from any checkpoint must reconstruct the same state. A destructive delete is an event whose effect depends on what was present when it ran, so replaying it against a divergent prefix produces a divergent result. A retraction tombstone is an additive event whose effect is the same under replay regardless of order against the facts it retracts, which is what makes replay deterministic. Provenance (PLRN-006) carries the lineage that replay reconstructs, and lineage requires that the events it records remain in the log.

3. The four operations without deletion

The critique’s four operations map onto existing primitives. The mapping is thick in two cases and thin in two, and the thin cases are where the architecture has work to do.

  1. Redaction. A fact that is wrong, or that must no longer be visible, is handled by a retraction tombstone. This is clean. The tombstone is an additive record asserting that a prior fact is retracted as of a given instant, and it satisfies audit and replay directly. The open question is propagation: a retraction tombstone has to propagate correctly through decay-based retrieval, so that a retracted fact is excluded from results rather than merely down-weighted, and the propagation has to hold across the same decay machinery that handles the other three operations rather than as a special case bolted onto it.
  2. Retirement. A fact that is not wrong, only now irrelevant, is handled by decay of its retrieval weight over time, called Mode 2 decay in Wunderblock’s bi-modal core. There is no explicit retire operation. A retired fact remains in the store, retains its provenance, and falls out of results because its decayed weight no longer clears the retrieval threshold, not because it was removed. Whether an explicit retirement operation is wanted is open. Decay handles the common case where relevance fades gradually; but an operator that knows a fact is now irrelevant has no way to say so directly and is left to wait for decay to do what an assertion could do at once.
  3. Supersession. A fact A replaced by a corrected or updated fact A' is handled today by asserting an inverse relation against A and inserting A'. This works but it is thin. The relationship between A and A' is reconstructable from the inverse relation, but it is not first-class, so a query that retrieves A' does not directly carry the fact that A' supersedes A. An explicit supersession edge is the candidate improvement: a typed edge from A' to A that the retrieval layer reads, so that surfacing A' can suppress or down-weight A as a single modeled relation rather than as two independent assertions a consumer has to correlate.
  4. Relaxation. A fact whose authority should weaken without disappearing is handled by per-fact authority weighting together with its decay rate. The fact stays, stays retrievable, and stays provenanced, and what changes is the authority the retrieval layer attaches to it. Relaxation is the operation that most clearly belongs at retrieval-time rather than in storage, because the fact is neither wrong nor irrelevant; only its standing has changed, and standing, in our construal, is a property of how the fact is scored and reported, not of whether it is kept.

4. Attention competition is a retrieval-time problem

The bi-modal core handles “stale competes for attention” at query time. A stale fact and a current fact may both match a probe. What wins is decided by recency- and decay-weighted scoring, not by having deleted the stale one. The stale fact does not compete on equal footing, because its decayed weight enters the ranking against the current fact’s, and the ranking is where the competition is resolved.

A recency limit at the result boundary already exists, which caps how far back results may reach. That limit is a coarse instrument. The deeper point is retrieval-time attention budgeting: the query has a finite budget of attention to allocate across matching facts, and the allocation should follow decay-weighted score, so that stale facts consume budget in proportion to their diminished weight rather than being either fully present or fully cut. Deleting the stale fact is one way to keep it from competing; but it is the wrong way, because it discards audit trail and replay determinism to solve a problem that scoring already solves.

5. Retrieval returns authority, not just relevance

Resolving the competition at query time is necessary but not sufficient. Even when a stale fact is correctly retrieved, surfaced because it genuinely matches the probe and is genuinely the best available answer, the consuming agent must know that it is stale. A hit that arrives as a bare match invites the agent to treat a low-authority fact as though it had full authority, which is the behavioral-authority failure zby’s critique properly names, reappearing one layer up from storage.

Lineage is already carried with retrieved facts, because provenance travels with the result (cf. our PLRN 006). It should also carry currency. A hit should arrive with a confidence-weighted authority signal, so the consumer receives not only that a fact matched but how current and how authoritative the match is. Retrieval returns authority, not just relevance. This is the claim that earns against the critique. The four lifecycle operations are realizable without deletion, the attention competition is resolved by scoring rather than erasure, and the behavioral-authority concern is answered by attaching authority to the hit, so that a stale fact that is correctly retrieved is also correctly discounted by the agent that consumes it.

A retraction tombstone is a logical operation. It records that a fact is retracted; the fact, its content, and its lineage remain in the store. This satisfies audit and replay, which is what append-only is for. However, it does not satisfy legal erasure.

A GDPR right-to-erasure request, or a contractual force-purge, requires that named records be physically removed, not just logically retracted. A tombstone leaves the content in place, so a tombstone is uncompliant, sensu stricto. Meeting that requirement needs a destructive-operator escape hatch that deliberately breaks the append-only invariant for named records. Breaking the invariant must itself be an audited, attested event (PLRN-006), so that the store can answer what was erased, by whom, under what authority, and when, even though it can no longer answer what the erased records contained. This is a genuine gap. The append-only default does not close it. Saying so is the point: append-only with retraction is the right default for everything audit and replay require, and legal erasure is the one case where the default has to be broken intentionally and the breaking has to be recorded just as intentionally.

7. What this is and is not

This is a defense of append-only memory with retraction against a specific and correct critique of attention competition. The defense concedes the failure the critique names and relocates the remedy from storage to retrieval, mapping redaction, retirement, supersession, and relaxation onto retraction tombstones, decay of retrieval weight, an inverse relation pending an explicit supersession edge, and per-fact authority weighting, and resolving attention competition through decay-weighted scoring with authority carried on every hit.

It is not a claim that append-only is sufficient by itself. Two of the four operations are thin today, namely supersession (which wants an explicit edge) and retirement (which may want an explicit operation alongside decay), and legal erasure requires a destructive escape hatch that the default does not provide. The honest position is that append-only with retraction is the correct default and that the exceptions are named, bounded, and audited rather than absent.

The critique answered here is zby, Designing Agent Memory Systems, which holds that lifecycle operations are non-negotiable for agent memory.

Append-only storage and event-sourcing lineage are long established. Immutable append-only logs and the CQRS / event-sourcing pattern reconstruct state by replaying recorded events, and the determinism argument in §2 is the standard event-sourcing argument applied to agent memory.

Bi-temporal and retraction semantics in databases give the precedent for retaining rather than overwriting. Datomic is accumulate-only and never overwrites a datom, recording retractions as additional facts, which is the database analogue of retraction tombstones. Datalog with negation and explicit retraction gives the logical semantics for asserting that a previously asserted fact no longer holds without removing the record of its having been asserted.

Decay and forgetting in memory systems, in which the retrieval weight of an item declines over time rather than the item being deleted, is the model behind Wunderblock Mode 2 decay and supplies the mechanism by which retirement and relaxation operate at retrieval time.

A note on method

Drafted in conversation with Claude Opus 4.8 (Anthropic) as structured interlocutor and prose editor. The framework, claims, and architectural commitments are mine.

Kendall Clark · k@pentad.ai
—Great Falls, Virginia
June 2026

Pentad Labs · pentad.ai