How We Rank Agent Context Management Options: The Criteria

Scoring

Options are ranked by their total score (Strong = 2, Partial = 1, Weak = 0, out of 12). Ties are broken by Scale, then by Retention of reasons, because those are the two that decide whether a long run stays on track.

Scores describe what an option does by default, as documented in its own docs, licence and published results. Where a fact isn’t published, we say so and point to where to check it rather than guess. A score is not a rating of the people who build a tool; it is whether the approach answers the question.

Fit matters as much as rank. An option that scores well can still be the wrong choice: a memory layer is overhead for a single short session, and each option’s page says where it is a poor fit.

Survival across sessions

Does what the agent was told still exist after the session, the machine or the sandbox is gone?

Context survives when a new session, on any machine, can reach what earlier sessions were told and decided without a person pasting it back in. Context that lives in one process, one laptop's home directory or one sandbox's disk does not survive; it only lasts until that thing is gone.

How to check it

Run a task, end the session and delete the working copy or sandbox.
Start a fresh session in a new checkout or container and ask it about a decision from the first run.
Note whether it can find the decision on its own, with nothing pasted in.

What each score means

Strong: A new session on another machine or in a new sandbox reaches earlier context on its own.
Partial: It survives only if someone commits it, or only on the same machine.
Weak: It ends with the session or the process.

How the options score

OpenCrawl	Strong	Transcripts are uploaded to a shared server, so a new machine or sandbox can search them.
General-purpose memory layers	Strong	Stored outside the session by design.
Notes and progress files	Partial	Survives only if it is committed and pushed before the session or sandbox ends.
Repo instruction files	Strong	In git, so every checkout and sandbox has them.
Session resume	Partial	Survives a restart on the same machine; not a new sandbox or another machine.
Compaction	Weak	Exists only inside the running session.
A bigger context window alone	Weak	Ends with the session.

Cross-tool reach

Can a different agent (Claude Code, Codex, Cursor and others) read it?

Teams rarely run one agent. Claude Code, Codex and Cursor each keep their own history in their own format, and none reads the others'. Cross-tool reach means context written while working in one agent can be read by another without conversion by hand.

How to check it

Make a decision in one agent's session.
Open a different agent on the same repo and ask about that decision.
Read the approach's docs for the list of agents it captures from and the list it serves to; they are often different lists.

What each score means

Strong: Several agents both write to it and read from it.
Partial: Several agents can read it, but only some write to it, or only through a manual step.
Weak: Only the agent that produced it can read it.

How the options score

OpenCrawl	Partial	Records Claude Code, Codex and Grok Bot sessions; any MCP client can search, but other agents are not captured yet.
General-purpose memory layers	Partial	Reachable from anything that calls the API or MCP server, but capture from each coding agent varies; check the integrations list.
Notes and progress files	Strong	Any agent that can read a file can read it.
Repo instruction files	Partial	AGENTS.md reaches many agents; tool-specific files (CLAUDE.md, Cursor rules) reach one. Check each agent's docs for which it reads.
Session resume	Weak	Only the agent that saved it can resume it.
Compaction	Weak	Each agent compacts in its own way, for itself.
A bigger context window alone	Weak	One model, one session.

Multi-agent sharing

Do agents working at the same time learn what the others were told while they work, not at merge time?

Several agents in one repo collide when each works from its own context: two change the same files and find out when the branches meet. Sharing means an agent can see, during its run, what other agents (yours or a teammate's) have been told and are changing.

How to check it

Start two agents on overlapping tasks in the same repo.
Tell one of them a constraint (a library choice, a file it owns).
See whether the other one learns it before it writes conflicting code, or only when the branches are merged.

What each score means

Strong: Agents see each other's context during the run.
Partial: Context is shared, but only after a commit, a merge or a manual handoff.
Weak: Each agent has only its own context.

How the options score

OpenCrawl	Strong	A message board with mentions and topic subscriptions delivered to agents while they run.
General-purpose memory layers	Partial	Agents can share one store, but awareness of what others are doing right now is not standard.
Notes and progress files	Weak	Shared through git, so other agents see it at merge time, often as a conflict.
Repo instruction files	Weak	Changes reach other agents only after they are committed and pulled.
Session resume	Weak	One agent, one session.
Compaction	Weak	Not shared.
A bigger context window alone	Weak	Not shared.

Retention of reasons

Does it keep why a decision was made, not just what was decided?

A summary keeps the plan and drops the conversation: which option was rejected, which constraint ruled it out, which ticket said so. Without the reasons, a later agent re-opens settled questions or undoes a choice it cannot see the point of.

How to check it

During a run, reject an option and say why ("not Modal: sandboxes can't reach the internet").
Later, from a fresh session, ask why that option was rejected.
Score it on whether the answer has the original reason in its original words, a paraphrase, or nothing.

What each score means

Strong: The original exchange is kept and can be retrieved word for word.
Partial: Some reasons are kept, as notes someone or something chose to write.
Weak: Only outcomes, or nothing, are kept.

How the options score

OpenCrawl	Strong	Keeps sessions word for word; search returns the original excerpts, not generated summaries.
General-purpose memory layers	Partial	Depends on what is stored; extracted facts and summaries tend to lose the why.
Notes and progress files	Partial	Keeps the reasons the agent bothered to write down, in its own paraphrase.
Repo instruction files	Weak	Rules, not the conversations behind them.
Session resume	Partial	Keeps the conversation until the session compacts, then only the summary.
Compaction	Weak	Reasons are what a summary drops first.
A bigger context window alone	Partial	Kept verbatim, but only while they fit and only if the model attends to them.

Scale

Does it still find the right thing when the history is far larger than a context window?

A long run, a team's month of sessions or a swarm produces far more text than any model can read at once. Scale is whether the approach still finds the right piece of history at millions of tokens, measured where possible on a public benchmark at large history sizes, such as BEAM's 1M and 10M token tiers.

How to check it

Look for published results on a public long-history benchmark (see /benchmarks), and check who ran them.
Prefer results at 1M and 10M tokens over results at sizes that fit in a context window.
If there are none, test it on your own history: ask for a detail from weeks ago.

What each score means

Strong: Published results hold up at 1M tokens or more, or retrieval does not depend on fitting history in a window.
Partial: It keeps working at scale, but by losing detail, or with no public evidence at large sizes.
Weak: It stops working once the history no longer fits in the window or in a file someone reads.

How the options score

OpenCrawl	Strong	83.69% on BEAM at 10M tokens in its own published runs (see /benchmarks for the method and caveats).
General-purpose memory layers	Partial	Retrieval-based, but published BEAM scores vary widely between products and fall at 10M tokens (see /benchmarks).
Notes and progress files	Weak	Read whole into context each time; it stops working once it is too long to read.
Repo instruction files	Weak	Must fit in every session's context, so they cannot grow with the project.
Session resume	Weak	Bounded by the context window; beyond it, it falls back to compaction.
Compaction	Partial	Lets a run continue indefinitely, by losing detail every time.
A bigger context window alone	Weak	Fails beyond the window; degrades before it (BEAM paper).

Openness

Can you read the code, run it yourself and take your data with you?

Agent context holds credentials, customer names and the reasons behind your architecture. Openness is whether the approach's code is under an open-source licence, whether you can run it on your own infrastructure, and whether the stored context is in a format you can read and move.

How to check it

Find the licence file in the repository and read its name (Apache-2.0, MIT, AGPL and so on).
Look for self-hosting docs and what they require.
Find out where the stored context lives and whether you can export it.

What each score means

Strong: Open-source licence and self-hostable, or plain files you own.
Partial: Readable data, but the logic that writes it is closed or vendor-controlled.
Weak: Closed, hosted only, or in a format you cannot read.

How the options score

OpenCrawl	Strong	Apache-2.0 server you can run in Docker on your own infrastructure; data in your own S3-compatible store.
General-purpose memory layers	Partial	Several are open source; hosted tiers and features differ. Check each licence.
Notes and progress files	Strong	Plain files you own.
Repo instruction files	Strong	Plain files you own.
Session resume	Partial	Usually readable local files, in a format the vendor controls.
Compaction	Weak	The summarising prompt and what it keeps are decided by the vendor.
A bigger context window alone	Weak	Depends entirely on the model provider.

What we leave out

We don’t score price, because the free options and the paid ones solve different problems, and prices change faster than reviews. Each option lists its cost where it is public. We don’t score ease of setup either: it matters, so each option’s limits say what setup takes.