← Analyses

Taxonomy archetypes

399 distinct fingerprints across 912 systems; top 12 cover 43.6% of the catalog. 93 recurring recipes (306 singletons in the long tail — 103 absorb as near-variants of recurring archetypes, leaving 203 genuinely bespoke designs).

What is an archetype?

Every record's seven taxonomy axes — storage · retrieval · persistence · update · unit · governance · conflict — collapse to a single fingerprint: the primary value on each axis. Systems with the same fingerprint share a "recipe". Recipes that recur many times are the memory landscape's archetypes; fingerprints with only one member are bespoke designs.

Methodology

We use the primary value on each axis, never the full array — multi-value axes legitimately exist (e.g. vector+graph storage) but hashing the full set would split the canonical recipe across half a dozen near-duplicates. Two systems are in the same archetype only when every one of the seven primary axis values matches; even one-axis disagreement (graph + traversal vs graph + injection) splits a recipe in two.

The "Hamming distance" clustering takes singletons (recipes with one member) and attaches them to the closest recurring archetype if they differ on a single axis. This collapses 103 of the 306 singletons into "near-variants" of existing recipes.

So what?

  • Top archetypes are well-travelled recipes — low risk, lots of prior art. If your design lands in one of them, assume you're competing on execution, not invention.
  • Singletons are either novel-and-interesting or one-off experiments. They become most actionable when paired with the influence-vs-adoption view: cited-but-not-adopted singletons are research bets the market hasn't followed.
  • Gap candidates are the white-space: every axis value is mainstream, but nobody has built this specific combination twice. The most interesting are the ones with a credible Hamming-1 neighbour you could "fork".
Tier
Section
#1

The no-taxonomy bucket (survey / event / meta record)

165 18.1%

Non-memory records — frameworks, surveys, infrastructure that the catalog tracks for context but has no memory taxonomy of its own.

  • storage n/a
  • retrieval n/a
  • persistence n/a
  • update n/a
  • unit n/a
  • governance n/a
  • conflict n/a
Median funding $30M 77 of 165
Median citations 35 3 of 165
Tier mix
T1·68T2·54T3·30T4·7T5·6
Sections
  • Training infrastructure 51
  • Agent frameworks (no first-party memory layer) 39
  • Evaluation & observability platforms 15
  • + 6 other sections
#2

The InstructLab (Red Hat / IBM) recipe

98 10.7%

None-trivial storage, none retrieval, session-only persistence; agent-controlled updates with opaque governance.

  • storage none-trivial
  • retrieval none
  • persistence session-only
  • update agent-controlled
  • unit turn
  • governance opaque
  • conflict out-of-scope
Median funding $46M 79 of 98
Median citations 0 of 98
Tier mix
T1·10T2·43T3·45
Sections
  • Use-case-specific agent harnesses 86
  • Multi-agent orchestration platforms 12
#3

The benchmark recipe (no memory of its own)

31 3.4%

Benchmarks — they measure other systems and have no memory taxonomy of their own.

  • storage n/a
  • retrieval n/a
  • persistence n/a
  • update read-only
  • unit n/a
  • governance n/a
  • conflict n/a
Median funding 0 of 31
Median citations 200 27 of 31
Exemplar LoCoMo T3
Tier mix
T2·2T3·23T4·6
Sections
  • Memory benchmarks & evaluation 30
  • Recent method papers — theorized, no distinct product 1
#4

The Cartesia Sonic recipe

17 1.9%

Kv storage, injection retrieval, session persistence; agent-controlled updates with opaque governance.

  • storage kv
  • retrieval injection
  • persistence session
  • update agent-controlled
  • unit turn
  • governance opaque
  • conflict stateless
Median funding $63M 14 of 17
Median citations 2.0k 1 of 17
Exemplar Cartesia Sonic T1
Tier mix
T1·11T2·5T3·1
Sections
  • Voice agent platforms 10
  • Computer-use & desktop agents 6
  • Agent IDEs & coding harnesses 1
1 near-variant — differ on one axis from the canonical recipe
#5

The Microsoft Phi-4 family recipe

13 1.4%

Parametric storage, parametric-recall retrieval, parametric-permanent persistence; read-only updates with opaque governance.

  • storage parametric
  • retrieval parametric-recall
  • persistence parametric-permanent
  • update read-only
  • unit weight
  • governance opaque
  • conflict n/a
Median funding $23.0B 13 of 13
Median citations 0 of 13
Tier mix
T1·9T2·4
Sections
  • Foundation models (substrate reference) 13
1 near-variant — differ on one axis from the canonical recipe
#6

The static-RAG recipe (read-only chunk retrieval)

12 1.3%

Research RAG: build the index, query it, never write back. The retrieval layer is treated as a static substrate.

  • storage vector
  • retrieval similarity
  • persistence long-term
  • update read-only
  • unit chunk
  • governance inspectable
  • conflict n/a
Median funding 0 of 12
Median citations 449 11 of 12
Exemplar Self-RAG T4
Tier mix
T3·5T4·7
Sections
  • Retrieval-as-memory hybrids 10
  • Recent method papers — theorized, no distinct product 1
  • Research / specialised systems 1
Example systems
2 near-variants — differ on one axis from the canonical recipe
  • RETRO governance: inspectable opaque
  • Memoro persistence: long-term session
#7

The Physical Intelligence (π) recipe

12 1.3%

Weight storage, parametric-recall retrieval, parametric-permanent persistence; agent-controlled updates with opaque governance.

  • storage weight
  • retrieval parametric-recall
  • persistence parametric-permanent
  • update agent-controlled
  • unit trajectory
  • governance opaque
  • conflict training-time
Median funding $350M 11 of 12
Median citations 1 3 of 12
Tier mix
T1·5T2·4T3·3
Sections
  • Robotics foundation models & agent stacks 12
#8

The CLAUDE.md recipe (file-backed agent memory)

11 1.2%

Files are the memory. Versioned in git, edited by humans and agents alike, no database in the loop.

  • storage file
  • retrieval injection
  • persistence cross-session
  • update overwrite
  • unit document
  • governance user-controllable
  • conflict manual
Median funding 0 of 11
Median citations 0 of 11
Exemplar AGENTS.md T1
Tier mix
T1·3T2·8
Sections
  • File-backed / editor paradigms 11
3 near-variants — differ on one axis from the canonical recipe
#9

The GitHub Copilot (Agent Mode) recipe

10 1.1%

File storage, injection retrieval, cross-session persistence; user-edit updates with inspectable governance.

  • storage file
  • retrieval injection
  • persistence cross-session
  • update user-edit
  • unit document
  • governance inspectable
  • conflict human-arbitrate
Median funding $10M 3 of 10
Median citations 0 of 10
Tier mix
T1·4T2·4T3·2
Sections
  • Agent IDEs & coding harnesses 10
#10

The Workato AI (Workflow Bot Builder) recipe

10 1.1%

None-trivial storage, none retrieval, ephemeral persistence; read-only updates with opaque governance.

  • storage none-trivial
  • retrieval none
  • persistence ephemeral
  • update read-only
  • unit file
  • governance opaque
  • conflict out-of-scope
Median funding $15M 9 of 10
Median citations 0 of 10
Tier mix
T1·2T2·6T3·2
Sections
  • AI sandbox & runtime environments 10
#11

The Qdrant recipe (vector chunk index)

10 1.1%

Index everything, retrieve by similarity, overwrite on re-ingest — the workhorse vector store.

  • storage vector
  • retrieval similarity
  • persistence long-term
  • update overwrite
  • unit chunk
  • governance inspectable
  • conflict overwrite
Median funding $38M 9 of 10
Median citations 0 of 10
Exemplar Qdrant T1
Tier mix
T1·8T2·2
Sections
  • Vector-database infrastructure 10
2 near-variants — differ on one axis from the canonical recipe
#12

The knowledge-graph recipe (fact graph, overwrite update)

9 1.0%

Facts as triples, queries as graph traversals — the knowledge-graph stack, optimised for relationship questions.

  • storage graph
  • retrieval graph-traversal
  • persistence long-term
  • update overwrite
  • unit fact
  • governance inspectable
  • conflict overwrite
Median funding $3M 7 of 9
Median citations 0 of 9
Tier mix
T1·4T2·5
Sections
  • Knowledge-graph platforms 9
2 near-variants — differ on one axis from the canonical recipe
  • Charisma.ai governance: inspectable user-controllable
  • Memgraph persistence: long-term session

The long tail

Beyond the top 12, 387 further fingerprints cover the remaining 56.4% of the catalog. 306 have only one member.

After Hamming-1 absorption: 103 singletons (33.7%) attach to a recurring archetype as near-variants, leaving 203 genuinely bespoke designs. The long tail collapses from 399 distinct shapes to roughly 296 (recurring archetypes + unattached singletons).

Threshold = 1 because that's the most legible boundary: "this system is the canonical recipe except for one axis" is easy to describe; distance ≥ 2 starts being a genuinely different shape. Anchor set = every archetype with ≥ 2 members, not just the top-12, so distance-1 singletons can attach to small-but-not-tiny recipes further down the list.

White-space candidates

Singletons whose individual axis values are common elsewhere — the building blocks exist in the literature, but this specific combination has never been built twice. The min axis popularity is the rarest axis value in the recipe, measured across the whole catalog; high min means every part is mainstream and what's missing is the assembly.

#1 Inworld AI min 137 · sum 1394

vector|similarity|long-term|extraction|episode|opaque|llm-arbitrate

The vector · similarity · long-term · extraction · episode · opaque · llm-arbitrate fingerprint has 0 members other than Inworld AI; every individual axis value is mainstream but this specific combination has not been built twice. Closest existing: Agent KB at Hamming-1 (differs on governance). Also nearby: AGiXT Adaptive Memory (governance), claude-supermemory (unit). A product filling this gap could be framed as "Agent KB with opaque on the governance axis".

#2 Generative Semantic Workspace (GSW) min 119 · sum 1356

vector|similarity|long-term|read-only|episode|inspectable|llm-arbitrate

The vector · similarity · long-term · read-only · episode · inspectable · llm-arbitrate fingerprint has 0 members other than Generative Semantic Workspace (GSW); every individual axis value is mainstream but this specific combination has not been built twice. Closest existing: Agent KB at Hamming-1 (differs on update). Also nearby: AGiXT Adaptive Memory (update), CrewAI Memory (update). A product filling this gap could be framed as "Agent KB with read-only on the update axis".

#3 Lyfe Agents min 119 · sum 1200

vector|similarity|session|extraction|episode|inspectable|llm-arbitrate

The vector · similarity · session · extraction · episode · inspectable · llm-arbitrate fingerprint has 0 members other than Lyfe Agents; every individual axis value is mainstream but this specific combination has not been built twice. Closest existing: Agent KB at Hamming-1 (differs on persistence). Also nearby: AGiXT Adaptive Memory (persistence), BrowserAgent (persistence). A product filling this gap could be framed as "Agent KB with session on the persistence axis".

#4 Krutrim Kruti (Ola) min 110 · sum 1211

vector|similarity|cross-session|extraction|episode|opaque|llm-arbitrate

The vector · similarity · cross-session · extraction · episode · opaque · llm-arbitrate fingerprint has 0 members other than Krutrim Kruti (Ola); every individual axis value is mainstream but this specific combination has not been built twice. Closest existing: BrowserAgent at Hamming-1 (differs on governance). Also nearby: ExpeL (governance), H²R (governance). A product filling this gap could be framed as "BrowserAgent with opaque on the governance axis".

#5 Botpress LLMz min 110 · sum 1164

vector|similarity|cross-session|extraction|document|inspectable|llm-arbitrate

The vector · similarity · cross-session · extraction · document · inspectable · llm-arbitrate fingerprint has 0 members other than Botpress LLMz; every individual axis value is mainstream but this specific combination has not been built twice. Closest existing: BrowserAgent at Hamming-1 (differs on unit). Also nearby: CREATOR (unit), ExpeL (unit). A product filling this gap could be framed as "BrowserAgent with document on the unit axis".

#6 MemGuide min 102 · sum 1156

vector|similarity|cross-session|extraction|fact|inspectable|llm-arbitrate

The vector · similarity · cross-session · extraction · fact · inspectable · llm-arbitrate fingerprint has 0 members other than MemGuide; every individual axis value is mainstream but this specific combination has not been built twice. Closest existing: Botpress LLMz at Hamming-1 (differs on unit). Also nearby: BrowserAgent (unit), ChatGPT Study Mode (governance). A product filling this gap could be framed as "Botpress LLMz with fact on the unit axis".

#7 Letta / MemGPT min 96 · sum 1346

kv|similarity|long-term|agent-controlled|episode|inspectable|llm-arbitrate

The kv · similarity · long-term · agent-controlled · episode · inspectable · llm-arbitrate fingerprint has 0 members other than Letta / MemGPT; every individual axis value is mainstream but this specific combination has not been built twice. Closest existing: FLEX at Hamming-1 (differs on storage). Also nearby: MemRL (storage), SCM (storage). A product filling this gap could be framed as "FLEX with kv on the storage axis".

  • FLEX differs on storage
  • MemRL differs on storage
  • SCM differs on storage
#8 Bedrock AgentCore (AWS) min 96 · sum 1331

kv|similarity|long-term|agent-controlled|fact|opaque|llm-arbitrate

The kv · similarity · long-term · agent-controlled · fact · opaque · llm-arbitrate fingerprint has 0 members other than Bedrock AgentCore (AWS); every individual axis value is mainstream but this specific combination has not been built twice. No other system in the catalog sits within one axis of this combination — the surrounding region is empty.