Research · Researchers and educators
Consolidation leakage: summaries and learned procedures in shared AI memory can expose restricted sources. Why record-level checks miss it, and which defenses to test.
Review pending. Report issues to editor@achiral.ai.
Consolidation Leakage in Team Memory
Research Preview
This essay is part of Achiral's research on Emergent Memory Systems for durable agents and team memory. It discusses mechanisms disclosed in U.S. provisional applications filed July 26 and November 9, 2026.
The architecture is proposed, not deployed. This essay reports no benchmark measurements. Its numerical examples are analytical calculations under stated assumptions. It extends the companion preprint, which covers the July baseline, with the November embodiments.

What is consolidation leakage in AI memory?
Consolidation leakage occurs when a summary, fact, or learned procedure exposes restricted information inherited from its sources, even though the requester is allowed to read the derived record. Checking that record's own access label can miss the information it carries from elsewhere.
A memory that learns doesn't only store what it is told. It summarizes conversations, turns episodes into facts, and turns repeated actions into procedures. Each step reads sources and writes something new. If the new memory receives only the label of the project where it was filed, a reader may be shown a summary built from notes they could never open. No restricted source record is returned to the reader. Its information can arrive anyway.
This matters in team memory: one memory that people and agents use together, with different permissions. A contractor and a support lead ask the same question. The memory has to distinguish what each may learn, including through derived content.
The failure mode is straightforward:
- A derived memory can carry information from restricted sources.
- A check based only on its destination label has no input describing those sources.
- The system needs a way to account for inherited information: conservative labels, separate consolidation, lineage-aware release, or another justified information-flow control.
There is a simpler design that may be enough for many teams: consolidate separately for each permission scope. A more elaborate lineage approach has to show that it preserves more useful knowledge at comparable leakage and cost. I don't know yet whether it does.
The companion architecture-and-analysis preprint develops the baseline projection and proposed evaluation. This essay also accounts for the later filed release-control mechanisms. Neither is a benchmark report.
How can a permitted procedure reveal restricted information?
Consider a hypothetical support team that resolves six synchronization tickets. The tickets are visible to the support project, contractors included. Internal customer-account notes are restricted. Five failures were fixed by refreshing an expired authorization token and replaying the sync. The sixth involved a suspended account.
A consolidation job might write a procedure: if the account is active and the token has expired, refresh the token and replay the sync. It stores the procedure in the support project. A contractor asks how to handle a sync failure and receives it.
The procedure may expose a condition learned from restricted notes. That is a possible information-flow violation even though the contractor may read the procedure. But the generic words "if the account is active" do not, by themselves, prove that a particular account—or any past account—was suspended. Whether the output reveals a protected fact depends on its wording and the reader's permitted background knowledge.
The access-control mistake begins when consolidation writes information into a record whose audience is broader than the information's allowed audience. Disclosure occurs when that information reaches the reader. The evaluation must test recovery of an actual protected fact, rather than treating every mention of an account condition as proof of a leak.
Why can record-level access control miss the leak?
Food labels are a useful comparison. You can't see the eggs in a cake, but the label still has to list them. "Baked in kitchen 3" would be accurate and useless to someone with an allergy.
A destination-only access label has the same limitation. It says where a memory was written. It doesn't necessarily describe what the memory was made from.
One remedy is to carry forward the most restrictive classification of the sources, using the policy's label lattice. This follows the information-flow tradition of Denning's lattice model. It can conservatively withhold a mixed-source procedure from contractors, including steps supported by tickets they may read.
A write-time label also becomes stale if it isn't updated when source permissions change. Propagating labels or recomputing authorization can address that problem, at an additional cost. Staleness is a limitation of an implementation, not an inherent property of labels.
Database security has long studied how permitted queries can reveal protected values through inference. Shared AI memory adds a setting where derivation can happen continuously, through language models, across agents and scopes. Its outputs become memories that are consolidated again. The research question concerns that evolving system, rather than a claim that information-flow security itself is new.
How do you measure leakage in a changing memory?
A probabilistic memory can produce different derived content from the same sources. Memories may gain activation when used, decay when unused, and be consolidated again. A language model may include a restricted detail on one run and leave it out on another.
A one-time inspection is therefore insufficient to characterize the leakage rate. The proposed test compares two memories:
- one built from the full event history;
- a counterfactual built without the restricted events, using matched construction settings.
An adversary receives the same requester's view in each case and tries to recover designated restricted facts. The difference in recovery accuracy estimates consolidation leakage for that adversary, target-fact distribution, and build process. Repeated builds and uncertainty intervals matter. A measured difference near zero does not prove confidentiality against every attacker.
This comparison also avoids counting facts that a reader could infer from permitted information alone. It asks what the restricted sources added to the reader's ability to recover the target fact.
What does lineage-closure projection do?
Lineage-closure projection builds a requester's view by checking a memory's ancestry as well as its own permissions. The July baseline uses an append-only record of derivation, retained when memories are edited, contradicted, or replaced.
Starting from selected memories, the projection checks nodes and applies an edge-disclosure policy. It may:
- Include an authorized edge and continue traversal.
- Tokenize a restricted endpoint, exposing only the relationship metadata permitted by policy.
- Omit a successor whose required ancestry cannot be disclosed.
- Mark permitted content as having incomplete lineage, when that marker is itself allowed.
Each cached view is tied to the requester's permissions and policy version. Authorization changes require invalidation or re-evaluation before another response is exposed. Different readers can receive different views of the same underlying graph.
Lineage has to capture the dependencies relevant to the decision. Recording a few convenient source citations is insufficient if an omitted source influenced the content or its visible state.
What does redaction still give away?
Hiding an ancestor does not sanitize its descendants. A token can conceal a restricted note while the procedure beside it still carries the note's information. Content release needs its own justification.
Withholding has a utility cost. Removing every mixed-source procedure can also remove steps supported by permitted evidence. Even then, missing lineage or other output channels can defeat the intended protection. Omission is a policy to test, not a blanket security guarantee.
Placeholders can reveal structure. In the baseline projection, tokens stable within a permission scope can link hidden sources across answers. Exposed placeholders may reveal existence, count, or class. The preprint calls this lineage disclosure. November's filed embodiments include request-unlinkable tokens and modes that suppress or coarsen this metadata. These choices still do not establish that adjacent derived text is safe.
The November filing also describes dependence estimation and a controller that selects a release action. At a high level, it can reconstruct from permitted inputs, retain supported claims, release a permitted abstraction, redact and mark incomplete, or withhold. Its calibration and robustness remain untested. A low dependence score is not a general proof of confidentiality.
Releasing procedure text is also separate from authorizing an action. Removing a restricted exception from the reader's view must not remove the exception from execution checks. The service carrying out the action still needs to enforce the applicable permissions and safety conditions.
This essay stays at the mechanism level. It does not publish production prompts, calibrated thresholds, canary facts, customer records, or an implementation recipe.
When should a learned procedure be allowed to act?
A procedure an agent can execute needs more than a readable access label. The July design describes promotion conditions for support count, success ratio, efficacy, safety, exceptions, rollback, and approval. Its representative values—at least five executions and a success ratio of at least 0.80—are illustrative, not calibrated production settings.
Small samples make the count-and-success subgate weaker than it looks. Under independent, identically distributed executions, a procedure that works half the time passes those two conditions 18.8% of the time when assessed on exactly five runs. This is not the probability of passing the whole gate: other conditions may prevent promotion.
Five successes in six runs give a 95% Wilson interval of approximately 0.44 to 0.97. That range includes a procedure that fails more often than it works. The interval is a statistical calculation, not an experimental result.
A stricter rule requires the lower bound to clear 0.80. At a single fixed look, the illustrated Wilson rule needs at least 16 successes in 16 trials, or more than 60 trials at a 90% observed success rate, to pass. It reduces false promotion under its assumptions; it cannot eliminate it. Correlation and repeated inspection change coverage and promotion probabilities. The IID calculations do not bound every dependent case.
Success also differs from efficacy. Good outcomes when a procedure ran do not establish that it caused them. November's filing specifies matched-comparison and independent-group estimators. They require defensible comparisons and dependence assumptions. Without randomization or justified identification assumptions, a difference in outcomes is evidence of association rather than proof of causation.
How would we test the proposed defenses?
The proposed evaluation uses synthetic organizational memory with planted restricted facts. That permits paired builds with and without the restricted events. It compares six policies:
- record-level filtering with two labelling rules;
- per-scope consolidation;
- lineage-closure projection with tokenization, omission, or an explicit release policy.
Per-scope consolidation is a meaningful baseline only when its inputs, caches, prompts, and learned artifacts obey the same permission separation. Merely putting outputs in separate namespaces is insufficient.
The evaluation measures restricted-fact recovery, lineage disclosure, recall of permitted knowledge, and cost. It should report repeated-build variation and uncertainty, and assess attacks beyond those used to tune the policy. Procedure-promotion tests should separately vary success rates, shared causes, and repeated decision-making.
Two outcomes would change my assessment. If destination-only filtering shows no excess recovery under a particular test, the hypothesized failure was not observed there; the test's coverage needs scrutiny. If per-scope consolidation matches lineage projection on leakage and usefulness at equal cost, the simpler design becomes the stronger practical choice. That second outcome is a real possibility.
Limits
- No benchmark measurements. The benefits, costs, and comparative performance remain hypotheses. Only the stated statistical examples have been calculated.
- Not a deployed system. This essay does not describe the behavior of any production memory, including Achiral's.
- Attacker-dependent evidence. Estimates depend on the adversary, target facts, and samples. A finite point estimate is not a guaranteed bound on other attacks.
- Synthetic data can mislead. Planted facts are cleaner than real knowledge, where permitted context may already reveal part of a restricted fact.
- Incomplete lineage limits protection. A policy cannot reliably reason about a dependency that was never recorded or reconstructed.
Related work
Collaborative Memory, by Rezazadeh and colleagues (2025), directly addresses multi-user, multi-agent memory with dynamic asymmetric permissions. Its fragments carry immutable provenance, and its policies support retrospective permission checks and transformed views. Those are close antecedents to the governance problem discussed here.
Lineage-Aware Memory Governance, by Sangaraju and Vissa (October 2026), attaches full derivation graphs to cached analytical results and gates retrieval on authorization for the source columns. Its guarantees are conditional on the stated lineage assumptions. This essay studies recursively consolidated memories, distinguishes hiding ancestry from sanitizing derived content, and proposes a comparative evaluation.
Agent memory systems such as Generative Agents, MemGPT, Zep, and A-MEM store, organize, and reflect on memories. Voyager and Agent Workflow Memory learn reusable procedures. Assumption-based truth maintenance tracks beliefs in different contexts. Training-data extraction and membership inference study how outputs reveal protected inputs.
This research builds on those fields. Its focus is a measurable failure mode in shared, continuously consolidating memory and the tradeoffs among defenses. Whether the proposed combination improves on per-scope consolidation remains an empirical question.
Publication and credits
Marvin Danig leads this research through his work at Achiral AI and is the sole named inventor on the associated U.S. provisional applications filed July 26 and November 9, 2026. Filing is a disclosure record, not validation of the architecture or a granted patent.
The companion preprint is Consolidation Leakage and Lineage-Closure Projection in Shared Multi-Agent Memory (2026). The preprint and this essay both distinguish the July baseline from the November embodiments.
Generative AI tools assisted drafting, revision, and analytical checks. Those tools are not authors or inventors. The illustration was generated with an image model from an Achiral prompt. No customer data or production evaluation results are presented here.
@misc{danig2026consolidationessay,
title = {Consolidation Leakage in Team Memory},
author = {Danig, Marvin},
year = {2026},
note = {Research preview, Achiral Research},
url = {https://achiral.ai/research/consolidation-leakage}
}Copyright 2026 Achiral Research. Patent rights reserved. No patent license is granted by publication of this article.