What the 243 current agentic-system analyses show about agent memory mechanisms (November 2026)

Type: types/note.md · Tags: agent-memory

The current accepted analyses connect memory to more than reference retrieval. Retained information also selects work, ranks candidates, gates admission and feeds later updates. This snapshot compares those mechanisms at one frozen evidence boundary. It asks how information enters memory, reaches later work, is revised and changes through recorded activity—not which product is best.

Four contrasts organize the findings:

• Admission: 221 of 225 code-grounded sets wire at least one automatic admission mechanism. Explicit human content decisions are wired in 63; another 55 support manual admission only through claims or affordances. Starting a workflow is not approval of each resulting memory item.
• Read-back: both requested retrieval and unsolicited context supply are extensively wired. Inferred relevance is not required for context supply: availability, chronology and identifiers also select what reaches a model. Selection must be distinguished from delivery and actual reliance.
• Curation: revision, compression and forgetting have many recorded implementations. Creating a claim absent from the inputs—synthesize in this comparison—has only 12 strong system witnesses. Calling a model or naming an operation “synthesis” does not establish semantic novelty.
• Trace-fed updates: the same trace_learning: "yes" classification includes access counters, retained mistake-prevention rules and trained weights. These change different material and reach different consumers. The shared classification is not an ordering of capability or benefit.

These are existence counts, not prevalence estimates. They count systems with a supported mechanism even when the rest of the inventory remains unresolved. They do not establish how often the mechanism runs or works.

Evidence boundary

• Mode and selection: current snapshot, all 243 accepted sets under kb/agentic-system-analyses/retained/, with no review filter. The successful builder supplies current membership and source-identity selection. Legacy reviews, archives, local runs, transfer advice and upstream sources supply no findings here.
• Inputs commit: bf2da04db03fb4cb60be52d1bb04e638a79eda79. It holds the retained members, contracts, reader and synthesis instruction. Relevant input paths were clean and unchanged from that commit before writing. Output paths are not evidence inputs.
• Matrix identity: the successful scripts/build_systems_matrix.py build, without a review filter, recorded SHA-256 19da70a6ca54b9c659dd9d1bf734e5f46885325d48f58b40685eccea083ce0bc. Its row identities, including manifest hashes, are the provenance used here. The public memory matrix and table are separately maintained outputs, not additional evidence.
• Population and revision: 225 code-grounded and 18 doc-grounded sets. All 243 profiles use comparison_version 2. Revision 2 measures human content-admission decisions, not generic caller identity; no revision-1 write-agency values are pooled with it.
• Boundary kinds: code-grounded sets comprise 193 whole-system, 27 complete-artifact/partial-loop and 5 subsystem-only boundaries. The doc-grounded split is 4, 12 and 2. Whole-system inclusion does not imply exhaustive inspection of every path.
• Source cutoffs: 2026-10-11 through 2026-11-09. Individual cutoffs and reviewed source revisions remain in the query ledger; this is not a simultaneous observation of deployed products.
• Lens: the ten axes of the memory profile contract. Units retain their scope, coverage, evidence basis and canonical records.
• Verification: local checking of selected retained members, profile units, canonical records and reconciliations. Selected directories were validated before reading members; the matrix builder validates population membership. Quantitative claims use the retained query output. No independent reviewer was commissioned.

How to read the evidence

A strong witness has basis wired, observed or causally supported in a code-grounded profile. wired means inspected code connects a mechanism, not that it ran successfully. Each system counts once per value, regardless of how many stores or routes support it. Both known and partial axes can support a positive witness. Weak-only means that a value is present only at claimed or afforded basis in that system; it is not added to the strong count.

Missing values and unresolved units are not negatives. An absent assessment needs bounded evidence. inapplicable, uninspected and not-determinable are different limitations, not observed absence. A strong witness for one unit does not upgrade a weaker unit with the same value.

Complete-value distributions remain withheld

Answering “what share is pull-only?” requires known axis coverage and strong evidence for every positive unit finding, not just the strongest witness for each union value. Few code-grounded sets qualify:

Axis	Eligible sets (of 225)
Storage substrate	36
Representational form	20
Read-back signal	17
Read-back direction	11
Behavioral authority	6
Write agency	2
Curation operations	1
Trace learning	1
Lineage; trace source	0

These assessed subsets do not stand in for the whole population. This snapshot therefore gives no complete-value distribution or whole-population percentage for any axis. Even a set with only files among its strong storage findings need not be a files-only system: weak or unresolved alternatives remain.

Wiring dominates; the observation exception is narrow

The 225 code-grounded profiles contain 9,379 wired, 849 claimed, 820 afforded and 2 observed findings. The 18 doc-grounded profiles separately contain 43 wired, 567 claimed and 22 afforded findings. Neither tier has a causally supported finding.

Both observed findings concern llm-wiki-newsroom's experiment artifacts: natural-language judgments and symbolic score/manifest fields at a later analysis consumer. The experiment-ledger object and judge-measurement route retain sampled labels and scores, not an independently authenticated production memory run. The reference labels are model judgments; hidden scorer/questions and missing originals limit replication and attribution. The observations do not establish deployed memory activation, improved answers or a reviewer-family effect.

Strong existence counts by axis

Every count below has denominator 225 code-grounded sets, revision 2. Counts overlap across values. Weak-only columns show selected frequent gaps; the ledger retains every value and every excluded run.

Axis	Axis assessment	Strong existence counts	Selected weak-only counts
Storage substrate	partial 168, known 56, not-determinable 1	files 204, in-memory 107, sqlite 77, vector 60, repo 45, graph 34, rdbms 34, service-object 16, kv 14, model-weights 7	repo 16, files 5, service-object 4
Representational form	partial 195, known 30	symbolic 221, natural-language 209, parametric 100	natural-language 14, parametric 3, symbolic 3
Lineage	partial 224, known 1	other-compiled 212, imported 200, trace-extracted 176, authored 131	authored 48, trace-extracted 23
Behavioral authority	partial 209, known 15, not-determinable 1	routing 211, knowledge 204, enforcement 190, ranking 183, validation 175, learning 151, instruction 116	instruction 45, learning 31
Write agency	partial 222, known 3	automatic 221, manual 63	manual 55, automatic 4
Curation operations	partial 223, known 2	evolve 184, consolidate 153, decay 134, invalidate 113, dedup 79, promote 59, synthesize 12	consolidate 42, evolve 31, dedup 24
Read-back direction	partial 196, known 29	pull 216, push 179	pull 9, push 7
Read-back signal	partial 166, not-determinable 21, known 19, uninspected 10, inapplicable 9	coarse 171, identifier 131, inferred-lexical 44, inferred-embedding 36, inferred-judgment 25	coarse 5, identifier 3, inferred-judgment 2
Trace learning	partial 187, not-determinable 36, known 1, uninspected 1	yes 160; local unit no in 6, system-union no in 1	yes 26
Trace source	partial 176, not-determinable 47, inapplicable 1, uninspected 1	session-logs 107, tool-traces 97, event-streams 46, trajectories 25	session-logs 24, tool-traces 16

Storage is not representation. Files have strong witnesses in 204 sets; 123 wire at least one database-type substrate (sqlite, rdbms, vector, graph or kv). A set can occur in both counts. In 36 sets files are the only strongly reported substrate, not necessarily the only substrate. Text, fixed-rule metadata and dense numerical representations can coexist within one store.

Authority is consumer-relative. Retained material supplies reference knowledge, but it also routes, ranks, validates and enforces. Those values do not imply that retained prose itself has veto power: symbolic state consumed by code and advice interpreted by a model are different paths. Instruction has 116 strong witnesses and 45 weak-only witnesses, illustrating the gap between requested standing guidance and a traced consumer.

Admission: human decisions, model proposals and software gates differ

Manual admission appears at some basis in 118 of 225 code-grounded sets. It has a strong witness in 63 and only weak support in 55. These are positive memberships, not a count of systems lacking human control. Of 1,250 write-agency units, 302 are not-determinable and 124 uninspected; a generic API caller often leaves the content decision unresolved.

CQ separates instructed approval from handler enforcement. Its shipped skill makes user approval a precondition of interactive propose. The approval claim classifies this as policy with afforded implementation. The proposal route checks structure and credentials, not approval; headless guidance permits model judgment under separate conditions. This does not establish a universal human gate, nor does a generic direct call establish automatic authorship.

cass_memory_system wires a real but scoped human decision. The trauma-admission route asks whether to add a displayed candidate batch to a permanent safety blocklist. That batch decision supports manual admission. Generic add/import callers remain unresolved, while other reflection and curation writes can be automatic. The human branch does not certify every future memory write.

pi-self-learning shows why invocation and admission must be separated. Automatic task-end reflection and /learning-now use the same proposal/parse/write route. A model receives selected branch activity and proposes mistakes/fixes; software accepts nonempty parsed lists and writes journals and derived rules. Manual invocation does not make each generated item manually admitted. Structural admission can reject malformed output without detecting a plausible false lesson. No prevention benefit is observed.

Read-back: unsolicited supply need not infer relevance

Pull has strong witnesses in 216 of 225 sets; push in 179. Here pull fulfills a consumer's request for retained material. Push selects and supplies retained material without that request. Delivery by code alone does not make a requested result push.

Among the 179 strong push sets, 71 also have a strong inferred selector (lexical, embedding or model judgment). The other 108 have strong coarse/identifier signals but no strong inferred finding. Partial coverage and weak alternatives prevent calling those 108 exclusively coarse systems. Only 4 sets establish a complete, strongly evidenced pull-only direction inventory. A relevance-ranked pull result is not evidence of relevance-based push.

Three routes expose distinct selection and consumption boundaries:

• Hindsight: a UserPromptSubmit hook derives a recall query from the prompt, filters scores and emits formatted retained text as additionalContext (RT-RTE-claude-hooks). The profile records embedding and cross-encoder-judgment selection for this route, while keeping unresolved retrieval arms and other integrations separate. Emitted context is not evidence of host reliance or complete transcript delivery.
• A-MEM: new-note ingestion supplies up to five embedding-selected retained neighbors to its own evolution model, which can revise links, tags and contexts (RT-RTE-neighbor-evolution). This push consumer maintains memory; it is not an answering agent. Partial mutation and uncertain snapshot refresh prevent a universal synchronized-read guarantee.
• pi-self-learning: the next-agent-start hook selects available core and optional dated files, plus recent process notes. Selection follows flags, chronology, counts and prefix caps, not the user's query (RT-RTE-context-injection). Long-term memory is named by path rather than automatically bundled. Later sections can disappear under character caps; process notes have a separate cap boundary. “Strict” consultation requests are not enforced reads.

These examples distinguish the selector, selected retained part, actual recipient and surviving payload. Storage availability does not guarantee that all reasons or conditions reach the next consumer. The focused read-back comparison examines that question in more detail.

Curation: changing retained content does not certify new knowledge

Revision (evolve) has strong witnesses in 184 of 225 sets; consolidation in 153, decay in 134 and invalidation in 113. There are 147 sets with three to seven strong operation values. These memberships describe recorded transformations, not exhaustive maintenance inventories.

The 12 strong synthesize witnesses are Agent-R, cass_memory_system, deja-vu, DocMason, eal-bench, Engraph, Kompl, LACP, Sparks, Swamp, VLM-wiki and Zikkaron. Another 8 sets support that value only weakly. Under the contract, synthesis creates a claim absent from its inputs; extraction, compression, index rebuilding and requested model behavior do not alone establish it.

In cass_memory_system, harmful feedback crossing a threshold can turn an eligible positive playbook rule into a new avoidance rule and deprecate the original. The curator-inversion record gives the code deduction Use cache → AVOID: cache. Marked harmful 3 times. This is policy synthesis, not accepted causal knowledge that avoiding cache works. Synthesis need not involve model inference.

Conversely, Signet's aggregate recall requests model-authored notes and can save eligible output by default without a maintenance pass (MEM-RTE-aggregate-recall). It links evidence, deduplicates or refreshes a note and attempts embedding; prior aggregate notes are excluded from aggregate searches to limit recursive reuse. But its save guard checks evidence eligibility and output surface form, not proposition-level entailment. Retention can survive embedding failure, and an unsaved result can still be returned. The profile supports deduplication and evolution without inferring semantic novelty from the “synthesis” prompt. Recall here can write memory; neither saving nor source linkage confers a truth warrant.

Revision and withdrawal also have consumer-specific horizons. Changing a current row does not by itself retract already delivered context or external copies. The focused correction comparison separates those propagation mechanisms; no population frequency follows from the examples selected here.

Trace learning: one classification, different durable changes

Trace learning in this profile means automatic trace-fed writes producing durable behavior-shaping artifacts or learned parameters. It does not require new knowledge or observed improvement. There are 160 strong "yes" witnesses among 225 code-grounded sets, plus 26 weak-only witnesses. Session logs, tool traces, events and trajectories supply different update inputs; their strong memberships are 107, 97, 46 and 25 respectively. These values can overlap and do not inventory every source in a partial boundary.

Four strong "yes" sets also wire model-weight storage: AgentFly, Amazon Science's SAGE, Self-Training-LLM and SkillRL. This co-occurrence does not classify every learning route in those systems as weight training. The retained records are needed to distinguish what changes:

• REM changes selection metadata. Retrieval asynchronously increments a durable count that can add a capped bonus to later rankings (MEM-OBJ-retrieval-count). This is use-fed feedback, not a checked outcome or model training. Its proposed semantic write-back is a different route: the shipped Go router lacks the accepting endpoints (RT-ABS-writeback-routes). A local learning "no" for that branch coexists with a system "yes" from the counter and other wired paths. Updates can fail silently; no ranking improvement is observed.
• pi-self-learning changes retained advice. Selected session activity becomes mistakes/fixes, journals and scored rules for later context. Its structural gate does not evaluate whether the lesson prevents recurrence. This route is distinct from REM's fixed numerical ranking feedback and from provider parameter training.
• SkillRL changes instructions and parameters on separate routes. Validation failures can produce new skills admitted to a training bank after structural checks. The proposer receives bounded failure excerpts and titles, not full bank principles. Separately, reward-driven policy updates change actor parameters. Bank persistence/resume and numerical checkpoint recovery are not one coupled guarantee. Neither route establishes better later actions here.
• A-MEM prevents an invalid inference from automatic revision. Neighbor evolution is automatic, but the original input is not established as an activity trace. Its trace-learning axis remains not-determinable. A bounded absence establishes that shipped paths do not update usage/history fields; it does not rule out caller-supplied experiences or all trace-fed writes.

Local negatives and a bounded system negative

Six sets retain a strong local unit "no": EchoWiki, ModularRSI, Phantom, REM, secure-llm-wiki and team-ai-workflows. Only team-ai-workflows has a known whole scoped learning inventory supporting the system-union "no". Its absence record covers all shipped corpus acquisition, lookup and access-maintenance paths: search reads files and prints results; synchronization changes access links; no automatic query/answer/feedback-to-retained-content route is shipped. Manual experience-derived authoring, external host/provider updates and adopter additions remain outside that negative.

The other five local negatives do not become system-wide absence. A "yes" elsewhere or unresolved included coverage prevents that inference. Likewise, the 36 not-determinable and one uninspected code-grounded learning axes are not negative observations.

The broad "yes" category cannot answer whether counters, summaries, rules or weight updates produce greater benefit. Selected records support the contrasts above. Population counts by those finer mechanisms would require a systematic classification with unresolved cases retained, not classification by missing prose mentions.

Doc-grounded tier: designs remain distinct from implementation

The 18 doc-grounded sets are AgeMem, aide2, arsumbris, claude-code-dynamic-workflows, context-constitution, From Agent Memory to Portable Skills, incremental-self-improvement, interview-doc-agent, IsaacCLupus_mnemosyn_spec, llm-context-base, llm-project-wiki, nimbalyst-wiki, pm-llm-wiki, Prove2Me_workspace, recall (parcha-skills), sig-releases, trajectory-informed-memory-generation and wikiskill. Thirteen carry claimed trace-learning "yes" findings, none a strong "yes". They do not enter the 225-set implementation counts.

A doc-grounded boundary can contain narrow code-inspected or executable configuration findings: the tier is the account's overall evidence boundary, not a basis assigned uniformly to every finding. The ledger retains all 43 wired labels in this tier without promoting the documented systems to code-grounded or assuming successful execution.

Two designs illustrate what the tier can contribute without that promotion:

• AgeMem describes conversation-to-store updates, replacement summaries consumed by later invocations and GRPO policy updates. The canonical store-maintenance, summary and policy-update records support these as claimed design from the captured paper, not inspected executor behavior. Summary fidelity, checkpoint lineage and production loading remain unresolved. A shared trajectory score does not identify the contribution of each memory action.
• Nimbalyst Wiki prescribes capturing decisions, rejected alternatives and reasons from sessions into subject pages, then reading a few full pages because reasoning surrounds the marks (RT-RTE-session-capture, RT-RTE-wiki-recall). Ordinary additions are agent-admitted without per-edit approval; destructive changes have separately scoped approval rules. CLI, host and service implementations are external to the frozen plugin. This supports a claimed trace-fed path and whole-page recall policy, not verified persistence, compliance or prevention of decision reversals.

Example provenance

Member links above name supporting sections and canonical IDs. The following identities come from the frozen matrix's manifest fields. Full source identities, reviewed revisions, member paths and hashes also remain in the ledger. Hashes identify ARTIFACT.yaml, not an upstream source commit. Each example uses the accepted member at the inputs commit.

Example	Run ID	Manifest SHA-256
A-MEM	AAS-2026-10-16-a-mem-sys-eff95cd80440-01	16fad56f4a88e833d4c0f5ffa129af99bca8a0383acbd6799d2dc99b28cd2f28
cass_memory_system	AAS-2026-10-13-cass-memory-system-30c18219cf68-01	2dcdb811124058cacf2dcde3b50b127dc826f2a8e603fb06ef92531b910c4729
CQ	AAS-2026-10-14-cq-8fb3dc41c334-01	8d2fcd66aa8707273424c810f716e9203b9f7c90be7296a5b1a10666c7d415aa
Hindsight	AAS-2026-10-14-hindsight-34c2ca821edf-01	3e9e67f75bff2a28a6ffc32ceb894615058008641f960da1e24bc9d70fb38579
llm-wiki-newsroom	AAS-2026-11-08-llm-wiki-newsroom-ea04687a5ff6-01	f16203ad8057cd9261819ac693dfc60e5eadd949f04f1afb80834032daf05bf0
pi-self-learning	AAS-2026-10-15-pi-self-learning-1a318610d4b1-01	903a6b625bf0d805a25d222dec241e857cf9bf8d0fe8677b009ff5e1832e0984
REM	AAS-2026-10-14-rem-a81fdc774400-01	0e45470aeb4b227b6deb7877ab75e9cc4db0199c9c6cf5aa590506243d3a3dee
Signet	AAS-2026-10-14-signetai-64f703d83203-01	4eb6367fe3884963d0c083b7c276bef866a089d4bcc8a1f0680bf1d7401bece9
SkillRL	AAS-2026-10-14-skillrl-0dba0e1c895e-01	9771fa25ae76ce88c6efdfc88fb4e126b1abe0bd3fac23ae510761d4acb361f4
team-ai-workflows	AAS-2026-11-09-team-ai-workflows-c00fb85f7ed1-01	3bbf81f1b11585b33fbc615ee2c7dcc994fee3f77782075805ecba7cd11b7c40
AgeMem	AAS-2026-10-12-2601-01885v1-32274e5bddc1-01	ec5ef488e3864432af6583e3410daeba596badd0839a85b58c257f4f7322306e
Nimbalyst Wiki	AAS-2026-11-08-nimbalyst-wiki-d9ff54186a76-01	0d48a874284207505fa604f5447fc3c1798179c57f4b9a10110f08cf90562a4f

Withheld conclusions and evidence gaps

• Prevalence, exclusivity and rankings: incomplete inventories prevent population shares or claims such as “only pull,” “no human control” or “files only” from missing positive witnesses. Named mechanisms establish examples, not prevalence of unstructured properties.
• Activation and benefit: wiring, delivery, model instructions, retained scores and experiment artifacts do not by themselves establish successful deployed use or improved task outcomes. No profile finding is causally supported.
• Semantic novelty and warrant: generated prose can compress, derive or invent. Without proposition-level input/output evidence, a model call does not settle which. Structural admission and provenance are not truth tests.
• Consumer completeness: truncation, optional context paths, separate stores and residual copies limit universal fidelity and withdrawal claims. No cross-system correction-coverage rate is inferred.
• Retained-set limits: partial units, opaque dependencies and verification limits remain operative. CQ and cass_memory_system retain coarse-part declaration issues outside the selected admission/inversion mechanisms; pi-self-learning retains a registration-count disagreement, not a disputed reflection route. Those issues are not repaired or silently generalized away here.
• Source-byte verification: selected validation logs include notices that some pinned source checkouts are unavailable for quotation checking. This synthesis checks committed accepted records, not upstream source bytes; it makes no fresh quotation-authentication claim or population-wide count of those notices.
• Change over time: this is one current snapshot. Counts do not establish growth, adoption trends or improvement relative to an older population. Historical outputs remain separate and are not pooled.
• Commonplace advice: design recommendations and transfer judgments are outside this public comparison.

Query ledger

Executable queries and dated outputs are retained in scripts/memory_landscape_queries/. refresh.py uses the existing unit decoder and strong-basis helper. Its outputs-2026-11-09/snapshot.json records revision and tier strata, field names, tests, basis filters, numerator, denominator, included runs and exclusions. Every assessment category is retained separately. evidence.py extracts selected committed record sections and profile units for semantic checking; it acquires no upstream evidence.

Query family	Membership or coverage test	Published use
Population	revision/tier split; boundary-kind and cutoff census	Evidence boundary
Axis existence	any strong unit witness for value; each system once	Ten-axis table
Weak-only	value present, no strong unit witness	Ten-axis table and admission contrast
Complete eligibility	known axis; every positive unit finding strong	Withheld distributions
Basis census	every finding, separately by tier	Wiring/observation boundary
Manual admission	manual at any basis versus strong	118 / 63 / 55 of 225
Push selection	strong push with any strong inferred signal; separately no strong inferred signal	71 / 108 of 179 push sets
Complete pull-only	known direction, strong set exactly pull, every finding strong	4 of 225
Storage	strong files set exactly {files}; any strong database-type substrate	36 reported strong-only files; 123 with database witness
Curation	any strong synthesize; three to seven strong operation values	12; 147 of 225
Trace learning	strong yes/units; local no retained apart from system unions; strong yes plus model-weights	160 yes; 6 local no; 1 system no; 4 weight co-occurrences

The query results come from the frozen matrix; smoke checks cover ledger structure and membership consistency. Historical outputs-2026-10-22/ remain separate; this snapshot makes no numerical change claim against them.