The Continuity Layer
Primer — The Landscape

The Landscape

Categories in the field this project maps. Entries are suppliers and reference points, not a ranking — most of them are solving a real, narrower problem than continuity, and the field needs all of them.

The Primer gives you the model; this page is where that model meets the systems actually shipping.
This map is early by design. The field is moving faster than the vocabulary — categories, statuses, and examples here will keep changing as the field does.
01

Memory systems

Active

Infrastructure that stores and retrieves facts about a user, agent, or session across time — the supplier layer most continuity systems are built on top of.

Why it matters

Storing facts is the part of continuity that's closest to solved. Most of the field's visible activity — and funding — sits here, which makes it easy to mistake "has memory" for "has continuity."

What to watch

Whether vendors start making project-level or decision-level claims (not just fact-recall claims) — that would be convergent-reinvention evidence, not a category collision.

Mem0 open-source memory layer for AI agents, persistent user/session facts.
Zep temporal knowledge-graph memory for agents (Graphiti); publishes its own evals.
Letta (formerly MemGPT) agent framework with built-in memory management and a public leaderboard for agentic memory.
ChatGPT Memory OpenAI's built-in cross-session memory feature for ChatGPT.
Graph-structured memory stores entities and relationships as the underlying representation (Zep's Graphiti is one public example).
02

Context engineering

Active

Prompt- and retrieval-management techniques — long-context handling, RAG, retrieval pipelines — that decide what a model sees in a given call.

Why it matters

This is where most "my AI forgot" complaints actually get patched today: bigger windows, better retrieval. It extends what a single session can hold, which is adjacent to continuity but not the same claim.

What to watch

The gap between advertised context length and usable context length (see RULER) — and whether retrieval-of-similar keeps getting sold as verification-of-current.

Long-context model research prompt- and context-management techniques that extend what a single session can hold.
Vector databases & RAG frameworks retrieval pipelines built for similarity search — retrieval of similar content, not reconstruction of current state.
03

Stateful agents

Emerging

Agent runtimes and frameworks that carry state across steps and tool calls within a task — distinct from raw memory storage and from multi-agent orchestration.

Why it matters

This is the layer that lets a single agent finish a multi-step job without re-deriving where it is. It's necessary for continuity but not sufficient — state-within-a-task isn't the same as understanding that survives a handoff to a different agent, tool, or person.

What to watch

Whether framework-level state management starts getting described in project-continuity terms, or stays scoped to single-agent task execution.

Agent runtimes & frameworks frameworks and runtimes that give agents state across steps and tool calls — adjacent to continuity claims, distinct from raw memory storage.
04

Agent infrastructure

Emerging

A real, active research field addressing attribution, identity, and interaction-shaping for AI agents — who acted, with what authority, leaving what trace.

Why it matters

It's the closest existing research to the questions this map tracks — who acted, with what authority, leaving what trace. An open question it raises but doesn't answer: what is accountability measured against? Audit trails need something to audit against, and whether that requires a project-level authoritative record is exactly the kind of underexplored question this map exists to track.

What to watch

Whether this research extends toward project-level records, or continues to operate at the level of individual agent actions.

Chan et al., "Infrastructure for AI Agents" (2025) attribution, interaction-shaping, and detection/remediation framing for agent infrastructure.
South et al., authenticated delegation (ICML 2025) research on verifying which agent acted under whose authority.
OpenID Foundation agentic-identity work standards-track effort on identity for AI agents.
05

Provenance & audit

Emerging

Work on tracing where a piece of AI-produced or AI-used information came from, and keeping an inspectable record of it.

Why it matters

Provenance is the evidentiary half of continuity: continuity says what changed and why; provenance says where the truth came from. Recent papers argue explicit provenance should be a requirement for responsible agentic systems, not an add-on.

What to watch

Whether provenance tooling stays scoped to individual outputs (a single answer's sources) or extends to a project's whole decision history.

"Responsible Agentic AI Requires Explicit Provenance" argues provenance tracking is a requirement for accountable agentic systems, not optional.
arXiv, "Right to History" proposes human sovereignty over an agent's computational history.
ADRs, decision logs the software-team ancestor of per-decision provenance — see Organizational memory, below.
06

Long-running workflows

Unclear

Orchestration and workflow systems built to survive interruption — checkpointing, retries, and resumption across long or multi-step processes.

Why it matters

Durable execution is a mature idea in distributed systems generally. Whether that maturity is being applied specifically to AI project continuity — versus just keeping a single long job alive — is an open, underexplored question this project is still mapping.

What to watch

Whether any workflow-orchestration product starts explicitly framing its value in terms of preserved project understanding, not just preserved execution state.

Durable-execution workflow engines the general category of systems built to checkpoint and resume long-running processes — continuity-adjacent, not continuity-specific by default.
07

MCP & tool continuity

Emerging

Protocol-level efforts to move context and tool access between different AI hosts and surfaces.

Why it matters

If context can move between hosts, that's a plumbing precondition for continuity — but the protocol carries context, not judgment about what that context means or why it changed.

What to watch

Whether MCP-adjacent tooling starts carrying decision lineage (not just tool schemas and context blobs) between hosts.

Model Context Protocol & related work protocol-level efforts to move context and tool access between different AI hosts and surfaces.
08

Personal AI memory

Active

Consumer-facing tools that build a personal memory or persona from an individual's own activity, notes, or data.

Why it matters

This is the individual-scale mirror of project continuity: useful, real, and adjacent — but scoped to one person's recall, not to a project understood the same way by multiple humans and agents over time.

What to watch

Whether personal-memory tools start being used as the substrate for multi-person or multi-agent project work, which would pull them toward project-level continuity.

Rewind personal recall tool that records and indexes a user's own device activity.
Personal.ai personal memory/AI-persona product built on a user's own data.
09

Organizational memory

Mature

Pre-AI institutional-memory practices — wikis, decision logs, architecture decision records (ADRs), audit trails, commit lineage — that already solved fragments of this problem for decades.

Why it matters

AI didn't create the continuity problem; it multiplied it. Decades of partial answers here are the strongest evidence the underlying problem is real, and ADRs specifically are the direct structural ancestor of what this project generalizes into a living system.

What to watch

Whether AI project tools start explicitly building on this lineage (citing ADRs, audit-trail practice) instead of treating memory as a green field.

ADRs, decision logs, provenance architecture decision records, decision logs, provenance tracking, and workflow-state tools — pre-AI institutional-memory practices this project draws structural parallels from.
10

Eval benchmarks

Active

Papers and leaderboards that attempt to measure memory, long-context, or continuity-adjacent properties directly.

Why it matters

Benchmarks are how the field checks its own claims. Most existing benchmarks measure recall or long-context handling; very few attempt to measure continuity properties (update-handling, disambiguation, model independence) directly — that gap is itself informative.

What to watch

Whether new benchmarks start testing update-handling and provenance together, rather than recall in isolation. See the Evidence Library for per-benchmark caveats and disputed results.

ATANT (v1.0, v1.1) Kenotic Labs' evaluation framework for continuity properties. See Evidence Library for caveats.
LOCOMO very long-term conversational memory benchmark, multi-session QA.
LongMemEval 500-question benchmark for long-term interactive memory in chat assistants.
BEAM synthetic long-conversation benchmark forcing retrieval-based memory over long-context stuffing.
MemoryBench benchmark for memory and continual learning in LLM systems via simulated user feedback.
RULER synthetic long-context suite measuring real usable context length.
11

Open-source projects

Watchlist

Publicly inspectable code implementing memory, continuity, or agent-state tooling — as distinct from company-run or closed products.

Why it matters

Open code is rarer than it should be in this field, and it's the fastest way for independent parties to verify a claim rather than take a vendor's word for it.

What to watch

Whether any of these projects attract independent contributors or replication attempts — single-maintainer repos are a different risk profile than actively co-maintained ones.

See Signals from the Field open-source entries (Continuity, ai-continuity-system, and others) are tracked there rather than duplicated here.
12

Papers

Watchlist

Academic and preprint research directly relevant to memory, continuity, provenance, or agent infrastructure.

Why it matters

Independent convergence across fields — ADRs in software, audit trails in accounting, provenance in science, "continuity of care" in medicine — is evidence the underlying abstraction is real, not invented for this project.

What to watch

New papers that name the project-level gap explicitly, rather than working one layer down at the agent-action level. See the Evidence Library for the specific papers this project's own claims are built on.

See Evidence Library this project's own sourced papers, with claims and caveats, are tracked there rather than duplicated here.
13

Commercial

Active

The business layer on top of memory, context, and agent-infrastructure technology — pricing, packaging, and go-to-market, as distinct from the underlying technique.

Why it matters

Watching what companies charge for (and what they bundle for free) is a signal of what the market currently believes is valuable — often narrower than what the research suggests is possible.

What to watch

Whether any commercial memory product starts pricing or packaging around project-level continuity, rather than per-seat or per-token fact storage.

Commercial memory-product tiers the paid tiers of the memory-systems vendors listed above (Mem0, Zep, Letta, and similar) — tracked here for business-model shape, not re-described technically.
14

Unsolved problems

Watchlist

Open, question-framed territory this project tracks rather than claims to have answered.

Why it matters

A map that shows where the abstraction breaks is more credible than one where it always works. This category exists so the site doesn't quietly overclaim.

What to watch

One underexplored question this project is tracking: whether long-running AI systems require project-level continuity in addition to agent-level infrastructure — and whether enough independent problems are converging that a category naturally wants to exist.

Where the abstraction breaks continuity requires time, understanding, and transfer between actors all at once — systems missing any one of the three (a raw database, an isolated program, a static archive) aren't a continuity problem. Robotics, largely perception-state rather than decision-state, is the clearest named counter-example.