From 642193efddd2b20560f945dbad8bf478f35b6a10 Mon Sep 17 00:00:00 2001
From: Isaac Kargar
- Lerim sits above agent traces, compiles the useful signal into cited context, and gives the next agent the operating memory it needs before work begins. + Lerim sits above agent traces, compiles useful signal into cited context and eval assets, and gives future agents the operating memory they need before work begins.
@@ -50,11 +50,11 @@ # Lerim -Lerim is a context compiler for AI agent workflows. +Lerim is a context compiler for repeated AI agent workflows. Agents leave traces everywhere: terminals, tools, tickets, code reviews, support cases, research runs. Most of that history is too noisy to reuse directly. -Lerim filters those traces into evidence-backed context records: the decisions, constraints, facts, preferences, and handoffs future agents should not have to rediscover. +Lerim filters those traces into evidence-backed context records and eval-ready workflow signal: the decisions, constraints, facts, preferences, corrections, and handoffs future agents should not have to rediscover. Instead of replaying raw traces or losing useful context between workflows, Lerim keeps: @@ -62,6 +62,7 @@ Instead of replaying raw traces or losing useful context between workflows, Leri - constraints - preferences - facts +- corrections - handoffs - evidence linked back to the source session @@ -70,7 +71,7 @@ Instead of replaying raw traces or losing useful context between workflows, Leri | Moment | Lerim does | Future agents get | | --- | --- | --- | | A completed agent run lands | Imports a source session from an adapter, MCP submit, or clean custom JSONL | A stable source boundary instead of a transcript paste | -| The trace is noisy | Compacts the run and filters for reusable decisions, constraints, facts, preferences, and handoffs | Durable context, not another log index | +| The trace is noisy | Compacts the run and filters for reusable decisions, constraints, facts, preferences, corrections, and handoffs | Durable context and eval-ready signal, not another log index | | Someone asks later | Retrieves relevant records and answers with citations back to stored evidence | A shorter start with less re-explaining | ## Quick Install @@ -101,7 +102,7 @@ lerim answer "What context should I know before working in this project?" ## Why Lerim -AI agents now triage tickets, investigate incidents, research markets, prepare handoffs, review policies, and change software. +AI agents now triage tickets, investigate incidents, research markets, prepare handoffs, review policies, analyze customers, and change software. Every run leaves a trace. Most traces are too long, too noisy, and too platform-specific for the next agent to reuse directly. @@ -111,16 +112,19 @@ Without a durable context layer: - constraints get rediscovered - preferences get ignored - every new session starts too close to zero +- useful corrections never become eval or training signal -Lerim fixes that by turning raw traces into reusable context records and making them queryable from agent tools and product workflows. +Lerim fixes that by turning raw traces into reusable context records, eval assets, and training-ready workflow signal that remain queryable from agent tools and product workflows. Lerim is meant for any trace-producing agent workflow. Today, native source adapters are strongest for coding agents, and documented custom-trace paths cover -support and incident workflows: +support and incident workflows. Coding is a proof-rich workflow pack, not the +whole product category: - coding agents: repo conventions, architecture decisions, setup facts, failed paths, test lessons, release handoffs - support operations: customer constraints, known fixes, failed fixes, escalation reasons, policy evidence, handoffs - operations and incidents: root causes, mitigations, rejected hypotheses, runbook gaps, incident handoffs, follow-up risks +- research, compliance, security, revenue, and other custom business agents: source trails, assumptions, approvals, rejected paths, policy facts, and workflow-specific handoffs when the source owner handles export, cleaning, and redaction ## Key Capabilities @@ -263,7 +267,19 @@ hardware/runtime metadata, and failure count. - Operations and incidents: documented custom-trace path; preserve root causes, mitigations, rejected hypotheses, runbook gaps, owner decisions, and follow-up risks. - Coding agents: retain architecture decisions, failed paths, repo conventions, setup facts, release handoffs, and constraints. -Research, revenue, security, and other verticals can use the same custom-trace path today when the user owns export, cleaning, and redaction. The first product wedge and strongest examples are coding plus support and incident operations. +Research, revenue, security, compliance, and other verticals can use the same custom-trace path today when the user owns export, cleaning, and redaction. The product wedge is one repeated private workflow with trace access, a workflow owner, privacy constraints, and measurable quality failure. Coding remains a strong proof workflow because the native adapters are mature, but the commercial company should be positioned around private agent improvement for enterprise workflows. + +## Enterprise Readiness To-Do List + +Use this list to keep the repo, website, and pitch aligned without turning the +open-source package into a closed enterprise product: + +- Keep open core useful: CLI, local runtime, MCP server, native adapters, custom trace import, context DB, docs, and benchmarks. +- Sell the production layer: Context Audits, private deployment, workflow evals, governance controls, managed integrations, retention, and enterprise support. +- Prove one workflow first: support escalation, incident/security ops, research intelligence, compliance review, or engineering automation. +- Measure improvement honestly: context reused, false memories rejected, eval pass rate, human acceptance, token budget saved, and repeated work reduced. +- Build training only after proof: approved traces, corrections, and eval assets can become SFT/RL data once the customer workflow and privacy boundary are clear. +- Keep coding agents as a proof pack, not the headline TAM/SAM/SOM story. ## Skill Updates @@ -292,7 +308,43 @@ lerim dashboard See [Skill Updates](docs/guides/skill-updates.md) for the dashboard workflow and [CLI: lerim skill](docs/cli/skill.md) for command details. -## Custom Agent Traces +## Custom & Non-Coding Agents + +Lerim is not only for coding agents. Support, incident/security operations, +research, compliance, revenue, and other custom business agents feed the same +compiler through clean JSONL traces and a signal profile that matches the workflow. + +
+
+
ad~}Z*Bg$FD?rX$c~`4o0FypL(Yg3r1ks?13vuW
zJPy&l9hkD-&`a)*!f5Gswo0j@04T-0_Xegl6liolAP-exkD_5wP1P7h&Q_RRC1Kcw
z-fQp+!nfI!^LQsE cH=xkNW(&pYoWF!S Uf&_^SGwT7i37JAl_-7UPmCv@*@hy
z9fOmIYAY~kb>Dh8_FqfZQ32c_1|oHMC5K5Pw8snC3=A<+KV=A>d*j!`Qcd`e0c#Jk
z%?0aRg4(%DIWAycFYsoe+4DH@AT*1+o@!Ihih_K@Jy~1RlNIYW9gE1@=k2E!3iO4nfc(FYK*AtKY)B2bf`-mPo=Pn>wnpwUL``jD5
zI}=}~<)upvEyO|K_uD>(_mvT;g5)!6&M#a8l0(r4;|!A3P}TelWsleVxJO;JHTX9>
zD@$-~hskguaBei+oKi%`mp$*}DkY@?Wj8%MyQ_yf%#!F)ldL{TmGbzcaCDTz(ubN5
zYRMIpXS%fU&s$ZY=Zg*Y2dS$U0LNpHceXVNFdgP7J^D4oG1v1wYgdmzp&1<2GMc6G
zB+*%zMby&ptp!weXsgDf(T*Y5l&F)8Qno46h2FR`c*~vWj>#571(pu*MjOzTrMP2V
zDyA}RBC7wtyLqcPpcT!c`E<*zQmdZSkShNaoGSoY=ww%8>topZCHkB{v18ahp?YN*-)v-b-tGHKBSg7w%?vT3b7I>vj79$k12GjY)7cxVqL*0
zG=eG($#c-hQVd5vODj||s!@mgD$&~|8y05>ZI+n4jl(Uo#a
zF=t~mk`#vJITwSCA8_~<@^p@JN19k8Qf7i2z_u2K=NE_1-WtCLkZbC5BJyxuue3of
zt^g5rW>9nXT);0bV6e0$uJ9b=cZqhg71TD>MhUKvLj?^J^1~1!@StKu
ovK(PTA@N!iR};$0(1svHbqA@(3lXVbSx?FmQoll93Y2FF)Ja-s+2K}JZfWF
z=w);*8;2$0h6GG{nsAAgPV0bOEVCNSF6-CS)Q}NoA`6JtSq4l`UY&qw6L>R)u`^vY
za%3AUT&)VR*P~(}_K+JhXAE>Oqe5E5?}>Jm0(C{1ZFO7FrSNoH9-Suqac)^Qw`F83
z!pk|m!kGa&1}b%W<1=m1K}Z1OqgN)1gV4~sGtkI|X-$C^|-3N9=Zp446q0$xN14exOgQo1I;yGAL;E
zh-n|gvuBpVl~7KBKZEfD6^vIumIbHW1rK}z6M+7$ec!o8%nUzk{A9#+bZ?HJBR@>j
z9+hIRuUi6^wIy(#%_*0J?sHQR_FP_;jBtisN-RgCy{BqO9>%rgIEcqim+pm{A!sod
zY9e(@5$`Y~$2n)F5|Ktc$1oT2m|F`c!f3tYN$X83RZW+9*d|j5og$5cPGA>CXGKN}
zK=oN6b*c-X1!}((S;f$
P(d%O
z^DBd&P-&7gWK~3&L5cS3VDV7I-Gp%Drq`Ec0^e2Bvy?qcGWT%(S>dnz`6nay`|oD(
zgzsn$7=wJ1nYmuMR{j0${q7f!@;sPj9%z&^QAWgC`Mh$idcSM$*xPcj^
z$*%a5`vU?XtKp;7YFM3h`|(bOuGhLt6`KlZg7%D3i!?4HyKwz2hY@;IaO!}RX!qOg
z$pPyFSQs~0m9I#ANYMBxBLuybG!73rC%J4w>NMt7(hscm{3)s4%YB