What nobody has solved
thecontinuitylayer.com/continuity/open-problemsNine problems, each with the obligation it belongs to, the sources that show it is open, and what would count as progress. This is the part of the site most likely to be wrong in a year, which is the point of writing it down with dates.
Checked 7 November 2026. Sources linked. Our own framing is labeled as ours.
The nine problems
01Govern
No benchmark scores governance
LoCoMo, LongMemEval and BEAM score recall; AMA-Bench scores recall over agent trajectories. The benchmarks that touch governance (GateMem, ForgetEval, MemSecBench, PAI-Bench) are each a single preprint with results nobody has reproduced, and no leaderboard reports any of them. A memory product can be at the top of every public table and never have been asked who may write to it.
Progress would look like: A reproduced benchmark that scores state mutation, access boundaries and deletion together, run by someone other than its authors, with vendors reporting on it.
Evidence: Always-On Agents: A Survey of Persistent Memory, Stat… · GateMem: Benchmarking Memory Governance in Multi-Prin… · Control-Plane Placement Shapes Forgetting: An Archite… · MemSecBench: Tracking Agent Memory Poisoning from Per…
02Carry
No shared format carries memory between vendors
Each memory system has its own schema. Export, where it exists, returns that schema; import, where it exists, expects it. The W3C community group chartered in June 2026 exists because of this, and had published nothing when checked. Until a baseline profile exists, portability is a one-vendor feature, and the right to data portability stops at a JSON file nobody else reads.
Progress would look like: A published interoperability profile with test vectors, and two independent memory systems that round-trip a user's memory through it.
Evidence: AI Agent Memory Interoperability Community Group, Cha… · The Sovereign AI Horizontal Memory (SAIHM) Protocol · GDPR, Article 17 (right to erasure) and Article 20 (r…
03Prove
Deletion has no receipt
A user who deletes a memory gets a confirmation, not evidence. Nothing says which rung of the ladder was reached, which derived artifacts were repaired, or which key was destroyed. The SAIHM draft and the ghost-vectors defense both propose signed receipts; no shipping assistant issues one.
Progress would look like: One major assistant or memory API issuing a signed, checkable deletion receipt that names the rung.
Evidence: Ghost Vectors: Soft-Deleted Embeddings Remain Reconst… · The Sovereign AI Horizontal Memory (SAIHM) Protocol · Deleted Is Not Forgotten: Deletion in Persistent LLM …
04Forget
Erasure is bounded by provenance, and provenance is incomplete
Cascade repair reaches zero exposure only when every dependency edge is known, and loses 17.7% of its coverage when 1% are missing. Provenance-guided replay matches a full reset only with conservative taint tracking. Most memory systems record no provenance at all at the write path, which is also why they can be poisoned. The two problems are one problem.
Progress would look like: Item-level provenance at the write path as a default in the open-source engines, and a measured bound on how much provenance a deletion guarantee needs.
Evidence: MemoRepair: Barrier-First Cascade Repair in Agentic M… · Forgetting Without Restarting: Execution-State Unlear… · From Untrusted Input to Trusted Memory: A Systematic …
05Carry
Identity is not carried
When the model under an assistant changes, the product calls it an upgrade and the user experiences a stranger. The research now has a vocabulary for what should survive (an identity contract, with lineage and governed updates) and a benchmark that shows agents recall identity facts and fail to compose them. No product exposes an identity contract a user can read, version or take with them.
Progress would look like: A deployed agent that survives a model swap with a versioned identity the owner can inspect, and a migration evaluation that was actually run.
Evidence: Runtime-Independent Persistent Agents: Preserving Ide… · Identity Is More Than Recall: A Benchmark for Persist… · AI Identity: Standards, Gaps, and Research Directions…
06Evolve
Conflicts have no rule
Two memories disagree, or two agents hold different versions of the same fact. Similarity retrieval cannot tell a contradiction from a duplicate (AUROC 0.59 in one 2026 test), so the stale value is served up to 40% of the time. Temporal graphs close the old edge when the new fact is clean; nobody has a rule for the messy case, and few systems are allowed to say “I cannot tell”.
Progress would look like: A stated, testable resolution rule (time, authority, source quality) with abstention, scored on a benchmark with real contradictions rather than templated ones.
Evidence: Temporal Validity in Retrieval Memory: Eliminating St… · LongMemEval: Benchmarking Chat Assistants on Long-Ter… · Zep: A Temporal Knowledge Graph Architecture for Agen…
07Forget
Consolidation resurrects what was deleted
Background processing that builds memory from history is now the consumer default (OpenAI's Dreaming, June 2026) and the research default (sleep-time compute). It is also a named resurrection route: consolidation write-back accounted for 12% of returned facts in the 2026 pilot, and checkpoint restores for 92%. The feature that makes memory feel effortless is the one that makes forgetting fail over time.
Progress would look like: Consolidation that reads tombstones: a deleted item's provenance excluded from every later summary, and a trajectory audit that shows it stayed excluded.
Evidence: Deleted Is Not Forgotten: Deletion in Persistent LLM … · ChatGPT memory with Dreaming · Sleep-time Compute: Beyond Inference Scaling at Test-…
08Forget
The law names outcomes, not methods
Article 17 says erase; Opinion 28/2024 says a model can hold personal data; neither says what erasure of a model, a summary or an embedding consists of. Research argues the common unlearning metrics reward hiding information at the output. A controller who deletes to rung 5 and a controller who crypto-shreds to rung 8 both say they complied, and no regulator has said which is right.
Progress would look like: Regulatory guidance that names a rung, or a court that does.
Evidence: GDPR, Article 17 (right to erasure) and Article 20 (r… · Opinion 28/2024 on certain data protection aspects re… · Machine Unlearning Doesn't Do the Thing You Think It … · Position: The Term “Machine Unlearning” Is Overused i…
09Prove
The words still mean different things
“Memory”, “context”, “state”, “knowledge” and “unlearning” are each used for several things, and 47 authors signed a survey in December 2025 saying the old categories no longer describe modern systems. This site's distinctions are a proposal; the field has not settled on any.
Progress would look like: Any shared vocabulary that two surveys and one standards body use the same way.
Evidence: Memory in the Age of AI Agents · Position: The Term “Machine Unlearning” Is Overused i… · Always-On Agents: A Survey of Persistent Memory, Stat…
What to build next
If the nine are read together, the order of work is clear and it is not the order the market is working in. The market is building better recall. The missing layer starts with provenance at the write path, because it is the one investment that pays into three problems at once: poisoning (an item that knows where it came from cannot be mistaken for the owner's own fact), erasure (an item that knows what was derived from it can be repaired), and accountability (a change that knows who made it can be audited). After that come deletion receipts, which turn a promise into evidence; then a portable profile, so a receipt from one vendor means something to another; then identity, so that what is carried has an owner.
This site's own practice is the smallest version of that order: a governed record with supersession and lineage, read first and written to last, which anyone can build with plain files. It reaches the carry, evolve and prove obligations and, honestly, stops at rung 5 of the ladder.
Our bets
Hypothesis Four hypotheses from the framework page, repeated here because this is where they will be tested.
H1 Forgetting is a provenance problem. A system can delete, to any rung of the ladder, only what it can trace; the completeness of the lineage graph bounds the completeness of erasure.
What would settle it: A study that achieves verified non-influence (rung 12) without item-level provenance, or one that shows provenance completeness does not predict residue.
H2 The five obligations are sufficient. Any continuity failure a user can name is a failure of carry, govern, evolve, prove or forget.
What would settle it: A reproducible failure class that none of the fifteen questions would have caught.
H3 Governance will become the differentiator. Recall scores on LoCoMo and LongMemEval are converging across memory systems; the next separation will come from state mutation, recovery and erasure, which almost no benchmark scores yet.
What would settle it: Two more years of leaderboards. If recall keeps separating products and governance scores stay unpublished, this was wrong.
H4 Portable memory will be adopted before verifiable memory. Export formats will ship years before signed receipts for deletion are common, because the first is a product feature and the second is a liability.
What would settle it: A major assistant shipping deletion receipts before a memory export.