Vocabulary drift The words the memory field uses to describe itself are not static:
"RAG" arrived in 2020, "agentic memory" became a phrase in 2024,
"knowledge-graph" peaked then receded. Tracking which terms enter,
spread, or fade — at the concept level rather than the system level —
catches recipe shifts the lineage view (which follows specific
systems) misses. A new vocabulary cluster is usually the first
signal that practitioners agree a new pattern matters.
27 curated memory terms tracked across
48 quarters · 1002 of 1002
records in scope after filters. ← back to analyses hub
How this view counts One record, one mention. If a term appears 12 times
in the same record's desc + claims we count
it once — we want "did this artefact adopt the vocabulary?", not
"how marketing-heavy was the copy?".Word-boundary, case-insensitive. "graph" matches
"graph" and "graphs" but not "graphical"; "RAG" matches "RAG-fusion".Bucketed by created date. Same parser as
the timeline view; records with year-only dates land in Q1 of that
year. Records with unparseable dates are excluded from the chart but
still contribute to co-occurrence and first-appearance counts.Curated 27-term list. Five thematic groups (cognitive
/ mechanism / agent-era / architectural / emerging). Extend via the
helper module + a DECISIONS.md note. Anything outside this list shows
up only as novel-vocab share in the ratio chart below — we
don't claim to track every term, just the ones we curate.Small-N badge. A growth % is flagged when fewer than
5 records sit on either side of the YoY ratio —
without this, "2 → 13" reads as "+550%" and dominates the eye. See
the "so what" callout for how to read flagged numbers.How to read it Fast-rising terms = emerging recipes. A term whose
cumulative arrival curve elbowed up in the last 4–6 quarters has
crossed from novelty into common vocabulary; the systems mentioning
it are usually converging on a new pattern. Open the drill panel
to see which.
Falling terms — two reads. Either the term is being
absorbed (e.g. "knowledge-graph" being subsumed by "graph-RAG"
where the same shape lives under a new label) or genuinely fading.
Compare against the cumulative chart: if cumulative is still rising,
the term is alive but quieter; if cumulative has plateaued for ≥4
quarters, it's actually receding.
Small-N rule. Percentages with the small-N badge are computed from fewer than 5 records on at least
one side of the ratio. Use the absolute "prior → latest" pair on the
same card; the percentage is mostly noise.
Tier: T1 T2 T3 T4 T5
Section: All sections (1002 records) AI sandbox & runtime environments Agent IDEs & coding harnesses Agent frameworks (no first-party memory layer) Agent identity & auth Browser-agent memory Claude Code memory mechanisms Coding-agent memory Computer-use & desktop agents Dedicated memory layers Durable execution & workflow engines Embedding & reranker services Enterprise-search adjacencies Evaluation & observability platforms File-backed / editor paradigms Foundation models (substrate reference) Framework-embedded memory Guardrails & AI security Inference platforms & gateways Knowledge-graph platforms MCP gateways & registries Memory benchmarks & evaluation Memory governance, privacy & safety Memory observability & monitoring Multi-agent orchestration platforms Personal AI / PKM / lifelogging memory Platform-provider memory Recent method papers — theorized, no distinct product Research / specialised systems Retrieval-as-memory hybrids Robotics foundation models & agent stacks Search platforms (non-memory) Theoretical / informal — ideas without a paper Training infrastructure Use-case-specific agent harnesses Vector-database infrastructure Vertical / domain-specific AI memory Voice agent platforms Voice-first / wearable AI memory Reset
First appearance by term The earliest year each tracked term shows up in any record's desc or claims within the current filter.
"Newcomers" arrived in 2024 or later — they are the
vocabulary candidates most worth watching.
Newcomers (0) · first seen ≥ 2024 Established (27) · first seen < 2024 1984 1995 2001 2001 2005 2005 2005 2005 2011 2016 2017 2017 2017 2018 2018 2019 2019 2019 2019 2020 2020 2020 2021 2021 2022 2023 2023 Total records 29 YoY (absolute) 13 → 6 (2025 → 2026) YoY (percent) -54% First seen 2017 Latest active 2026-Q2 Top mentions: Anthropic Claude (foundation models) ,Lindy AI Memory ,LinkedIn Cognitive Memory Agent
Total records 19 YoY (absolute) 10 → 1 (2025 → 2026) YoY (percent) -90% First seen 2017 Latest active 2026-Q2 Top mentions: Botpress LLMz ,Lovable ,OpenAI Agents SDK Memory
Total records 13 YoY (absolute) 5 → 3 (2025 → 2026) YoY (percent) -40% First seen 2020 Latest active 2026-Q2 Top mentions: Anthropic Claude (foundation models) ,OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) ,ICLR 2026 MemAgents Workshop
Total records 62 YoY (absolute) 19 → 6 (2025 → 2026) YoY (percent) -68% First seen 1984 Latest active 2026-Q2 Top mentions: Cline (framework) ,Model Context Protocol (MCP spec) ,Claude Code (Anthropic)
Total records 8 YoY (absolute) 2 → 2 (2025 → 2026) YoY (percent) 0% small-N First seen 2023 Latest active 2026-Q1 Top mentions: O-Mem ,ELDER ,SimpleMem
Total records 65 YoY (absolute) 19 → 5 (2025 → 2026) YoY (percent) -74% First seen 2005 Latest active 2026-Q2 Top mentions: Haystack (framework) ,OpenAI Embeddings ,Coveo
Term co-occurrence (top 15) How often two terms appear in the same record's desc + claims. Darker = more co-mentions. Edges with count
≥ 4 form the cluster signal. Click a row label to
drill into that term.
retrieval graph MCP RAG agentic semantic vector persistent episodic hierarchical attention working procedural latent KV retrieval retrieval ∩ retrieval = 105 (self) 105 retrieval ∩ graph = 18 18 retrieval ∩ MCP = 9 9 retrieval ∩ RAG = 26 26 retrieval ∩ agentic = 4 4 retrieval ∩ semantic = 18 18 retrieval ∩ vector = 18 18 retrieval ∩ persistent = 2 2 retrieval ∩ episodic = 7 7 retrieval ∩ hierarchical = 4 4 retrieval ∩ attention = 4 4 retrieval ∩ working = 1 1 retrieval ∩ procedural = 1 1 retrieval ∩ latent = 1 1 retrieval ∩ KV = 4 4 graph graph ∩ retrieval = 18 18 graph ∩ graph = 82 (self) 82 graph ∩ MCP = 10 10 graph ∩ RAG = 11 11 graph ∩ agentic = 7 7 graph ∩ semantic = 11 11 graph ∩ vector = 16 16 graph ∩ persistent = 4 4 graph ∩ episodic = 5 5 graph ∩ hierarchical = 5 5 graph ∩ attention = 0 graph ∩ working = 0 graph ∩ procedural = 0 graph ∩ latent = 0 graph ∩ KV = 2 2 MCP MCP ∩ retrieval = 9 9 MCP ∩ graph = 10 10 MCP ∩ MCP = 82 (self) 82 MCP ∩ RAG = 3 3 MCP ∩ agentic = 8 8 MCP ∩ semantic = 7 7 MCP ∩ vector = 5 5 MCP ∩ persistent = 5 5 MCP ∩ episodic = 1 1 MCP ∩ hierarchical = 0 MCP ∩ attention = 0 MCP ∩ working = 2 2 MCP ∩ procedural = 0 MCP ∩ latent = 0 MCP ∩ KV = 0 RAG RAG ∩ retrieval = 26 26 RAG ∩ graph = 11 11 RAG ∩ MCP = 3 3 RAG ∩ RAG = 67 (self) 67 RAG ∩ agentic = 7 7 RAG ∩ semantic = 10 10 RAG ∩ vector = 9 9 RAG ∩ persistent = 2 2 RAG ∩ episodic = 4 4 RAG ∩ hierarchical = 2 2 RAG ∩ attention = 2 2 RAG ∩ working = 2 2 RAG ∩ procedural = 1 1 RAG ∩ latent = 0 RAG ∩ KV = 3 3 agentic agentic ∩ retrieval = 4 4 agentic ∩ graph = 7 7 agentic ∩ MCP = 8 8 agentic ∩ RAG = 7 7 agentic ∩ agentic = 64 (self) 64 agentic ∩ semantic = 2 2 agentic ∩ vector = 5 5 agentic ∩ persistent = 1 1 agentic ∩ episodic = 0 agentic ∩ hierarchical = 2 2 agentic ∩ attention = 0 agentic ∩ working = 1 1 agentic ∩ procedural = 0 agentic ∩ latent = 0 agentic ∩ KV = 1 1 semantic semantic ∩ retrieval = 18 18 semantic ∩ graph = 11 11 semantic ∩ MCP = 7 7 semantic ∩ RAG = 10 10 semantic ∩ agentic = 2 2 semantic ∩ semantic = 60 (self) 60 semantic ∩ vector = 8 8 semantic ∩ persistent = 6 6 semantic ∩ episodic = 13 13 semantic ∩ hierarchical = 3 3 semantic ∩ attention = 1 1 semantic ∩ working = 4 4 semantic ∩ procedural = 4 4 semantic ∩ latent = 0 semantic ∩ KV = 0 vector vector ∩ retrieval = 18 18 vector ∩ graph = 16 16 vector ∩ MCP = 5 5 vector ∩ RAG = 9 9 vector ∩ agentic = 5 5 vector ∩ semantic = 8 8 vector ∩ vector = 52 (self) 52 vector ∩ persistent = 2 2 vector ∩ episodic = 2 2 vector ∩ hierarchical = 1 1 vector ∩ attention = 0 vector ∩ working = 3 3 vector ∩ procedural = 1 1 vector ∩ latent = 1 1 vector ∩ KV = 1 1 persistent persistent ∩ retrieval = 2 2 persistent ∩ graph = 4 4 persistent ∩ MCP = 5 5 persistent ∩ RAG = 2 2 persistent ∩ agentic = 1 1 persistent ∩ semantic = 6 6 persistent ∩ vector = 2 2 persistent ∩ persistent = 44 (self) 44 persistent ∩ episodic = 0 persistent ∩ hierarchical = 0 persistent ∩ attention = 0 persistent ∩ working = 2 2 persistent ∩ procedural = 0 persistent ∩ latent = 0 persistent ∩ KV = 1 1 episodic episodic ∩ retrieval = 7 7 episodic ∩ graph = 5 5 episodic ∩ MCP = 1 1 episodic ∩ RAG = 4 4 episodic ∩ agentic = 0 episodic ∩ semantic = 13 13 episodic ∩ vector = 2 2 episodic ∩ persistent = 0 episodic ∩ episodic = 29 (self) 29 episodic ∩ hierarchical = 1 1 episodic ∩ attention = 0 episodic ∩ working = 4 4 episodic ∩ procedural = 4 4 episodic ∩ latent = 0 episodic ∩ KV = 1 1 hierarchical hierarchical ∩ retrieval = 4 4 hierarchical ∩ graph = 5 5 hierarchical ∩ MCP = 0 hierarchical ∩ RAG = 2 2 hierarchical ∩ agentic = 2 2 hierarchical ∩ semantic = 3 3 hierarchical ∩ vector = 1 1 hierarchical ∩ persistent = 0 hierarchical ∩ episodic = 1 1 hierarchical ∩ hierarchical = 26 (self) 26 hierarchical ∩ attention = 0 hierarchical ∩ working = 1 1 hierarchical ∩ procedural = 0 hierarchical ∩ latent = 1 1 hierarchical ∩ KV = 0 attention attention ∩ retrieval = 4 4 attention ∩ graph = 0 attention ∩ MCP = 0 attention ∩ RAG = 2 2 attention ∩ agentic = 0 attention ∩ semantic = 1 1 attention ∩ vector = 0 attention ∩ persistent = 0 attention ∩ episodic = 0 attention ∩ hierarchical = 0 attention ∩ attention = 23 (self) 23 attention ∩ working = 1 1 attention ∩ procedural = 0 attention ∩ latent = 2 2 attention ∩ KV = 4 4 working working ∩ retrieval = 1 1 working ∩ graph = 0 working ∩ MCP = 2 2 working ∩ RAG = 2 2 working ∩ agentic = 1 1 working ∩ semantic = 4 4 working ∩ vector = 3 3 working ∩ persistent = 2 2 working ∩ episodic = 4 4 working ∩ hierarchical = 1 1 working ∩ attention = 1 1 working ∩ working = 19 (self) 19 working ∩ procedural = 2 2 working ∩ latent = 3 3 working ∩ KV = 2 2 procedural procedural ∩ retrieval = 1 1 procedural ∩ graph = 0 procedural ∩ MCP = 0 procedural ∩ RAG = 1 1 procedural ∩ agentic = 0 procedural ∩ semantic = 4 4 procedural ∩ vector = 1 1 procedural ∩ persistent = 0 procedural ∩ episodic = 4 4 procedural ∩ hierarchical = 0 procedural ∩ attention = 0 procedural ∩ working = 2 2 procedural ∩ procedural = 16 (self) 16 procedural ∩ latent = 2 2 procedural ∩ KV = 0 latent latent ∩ retrieval = 1 1 latent ∩ graph = 0 latent ∩ MCP = 0 latent ∩ RAG = 0 latent ∩ agentic = 0 latent ∩ semantic = 0 latent ∩ vector = 1 1 latent ∩ persistent = 0 latent ∩ episodic = 0 latent ∩ hierarchical = 1 1 latent ∩ attention = 2 2 latent ∩ working = 3 3 latent ∩ procedural = 2 2 latent ∩ latent = 14 (self) 14 latent ∩ KV = 0 KV KV ∩ retrieval = 4 4 KV ∩ graph = 2 2 KV ∩ MCP = 0 KV ∩ RAG = 3 3 KV ∩ agentic = 1 1 KV ∩ semantic = 0 KV ∩ vector = 1 1 KV ∩ persistent = 1 1 KV ∩ episodic = 1 1 KV ∩ hierarchical = 0 KV ∩ attention = 4 4 KV ∩ working = 2 2 KV ∩ procedural = 0 KV ∩ latent = 0 KV ∩ KV = 14 (self) 14
Threshold ≥ 4 shown with orange border.
Max off-diagonal: 26.
Stable vs novel vocabulary per year For each year, the share of new records whose desc + claims mention at least one of our 27
curated terms (stable) vs records using none of them (novel /
drifting). A rising novel share = the field's vocabulary is
outrunning our curated list. Bar width is the record count for the
year (normalised to the largest); rising novel slice = vocabulary
churn.
Year Records Stable Novel Stable share Mix 1984 1 1 0 100% 1995 2 1 1 50% 1999 1 0 1 0% 2001 2 2 0 100% 2003 1 1 0 100% 2005 2 2 0 100% 2006 1 0 1 0% 2007 3 1 2 33% 2008 2 2 0 100% 2009 2 0 2 0% 2010 3 2 1 67% 2011 3 1 2 33% 2012 5 3 2 60% 2013 5 3 2 60% 2014 12 4 8 33% 2015 15 2 13 13% 2016 11 4 7 36% 2017 18 4 14 22% 2018 30 9 21 30% 2019 33 12 21 36% 2020 34 12 22 35% 2021 37 12 25 32% 2022 60 24 36 40% 2023 196 80 116 41% 2024 219 109 110 50% 2025 196 126 70 64% 2026 59 45 14 76%