Your agent is brilliant. Its knowledge is frozen.
Claude's training ended months ago. The framework you use shipped a new API last week. WebBrief closes the gap: one call returns a cited, ~500-token brief — briefs, not pages — and your agent's context stays intact for the actual work.
$ npx webbrief login
Agents can self-onboard from agents.md.
❯ Has anything replaced plain vector stores for agent memory this year?
bottom line
Partly — pure vector retrieval is now the fallback, not the default.
Production agent stacks have largely moved to layered memory: a structured working set, episodic summaries, and vector search only as recall of last resort [1].
Two widely-cited postmortems this spring attribute agent drift to unfiltered vector recall; both teams replaced it with scoped, time-decayed retrieval [2].
Sources disagree on graph memory: one benchmark shows wins on multi-hop tasks, maintainers of two major frameworks call it premature [3].
9 sources cited · all dated · as of today
sample recording — final numbers ship from the public evals repo
You already know the workaround. It's why your sessions die.
WebSearch returns links. WebFetch dumps 15,000 tokens of one page into your context. Comparing five sources means 50,000 tokens of navigation chrome, cookie banners, and marketing soup — and the answer still isn't settled.
So you either let the agent code from stale memory, or you sacrifice the session to find out what's true.
The research happens on our servers. The answer lands in your context.
Fresh
sources: live · dated
Multi-source and live: search engines, GitHub issues and discussions, Reddit, YouTube transcripts, papers. Dated results, always.
Cited
claims: attributed
Every claim has a source and a date. When sources disagree, the brief says so instead of picking a side silently.
Bounded
cap: ~500 tokens
~500 tokens, guaranteed. Your agent reads the answer, not the internet — context stays intact for the actual work.
The full procedure — triangulation, citation rules, the quality rubric — is public: methodology.
The same answer, at 1–2% of the tokens.
Parity judged blind against the raw-fetch workflow.
| Task type | WebSearch + WebFetch | WebBrief | Answer parity |
|---|---|---|---|
| Library upgrade check | 38,400 | 610 | same answer |
| Compare 3 frameworks | 52,100 | 780 | same answer |
| Debug a fresh API change | 24,900 | 540 | same answer |
| State of a technique | 61,300 | 820 | brief more current |
Table 1 — Tokens consumed per research task. Sample data; launch numbers are measured and reproducible from the public evals repo.
Your agent knows a million ways to build this. It defaults to average.
A brief in context aims generation at the frontier of a topic instead of its statistical center — what's current, what's contested, what actually works now. Knowledge gaps close as models improve; direction is needed forever.
Before you build— the current pattern, not last year's. Before you choose — what maintainers and postmortems actually say. Before you trust the model — whether that API is still shaped the way it remembers.
- 4,812
- briefs served
- 512
- avg tokens per brief
- 31,207
- sources cited
- 5,000/mo
- run by the founder
Static sample values — these counters go live from the API before launch.
Pricing
Free
$0
25 briefs/mo
Pro
$19/mo
500 briefs/mo
Max
$49/mo
2,000 briefs/mo
Deep briefs meter higher · credit packs for overage · never unlimited · full pricing, incl. Power tier