#Memory solutions
charred nestBOT
solid lanternBOT
Memory.md + a daily log is a good start, but it’s not enough by itself for an “operator-run company” agent to stay coherent over months. What you want is a 3-layer memory system:
- Constitution (how the agent behaves)
- Knowledge base (what the company is)
- Records (what happened + why decisions were made)
If you build only (3), the agent “remembers events” but never stabilizes into a business.
1) Constitution (AGENTS.md / SYSTEM prompt equivalent)
Put the rules of the business here, not random facts.
Include:
- Mission + constraints: what the company does / does not do
- Risk & approvals: “Ask before spending money, signing up for paid tools, emailing customers, touching legal/tax, changing pricing, deleting data”
- Operating cadence: daily/weekly/monthly loops
- Tool policy: what tools it can use, and when it must ask
- Definition of done: what “progress” means (revenue, leads, shipped features)
This is what prevents “agent drift”.
2) Knowledge base (durable, structured docs)
This is the missing piece for most people. Treat it like a tiny internal wiki.
Create separate files (or sections) for:
- COMPANY.md – one-paragraph description + positioning
- OFFER.md – what you sell, deliverables, boundaries
- ICP.md – ideal customer profile, disqualifiers
- PRICING.md – current pricing + rationale
- PRODUCT.md – features, roadmap, “not doing” list
- SOPS/ – repeatable playbooks (sales outreach, onboarding, support, shipping)
- BRAND.md – tone, voice, examples of good/bad messaging
This is what lets the agent onboard itself again after resets.
3) Records (memory.md + logs + decisions)
memory.md (durable facts only)
Use this for things that should stay true until explicitly changed:
- People + preferences (important customers, partners)
- Current strategy choice (e.g., “B2B service first, product later”)
- Current offer + pricing
- Active constraints (budget limits, time limits)
- Current KPIs / north star metric
Keep it short. If it grows without bound, retrieval gets noisy and the agent starts hallucinating relevance.
Daily log (episodic)
Daily logs are great, but only if you distill them.
Daily template (example):
- Wins shipped today
- Pipeline updates (new leads, follow-ups, replies)
- Revenue events (invoices sent/paid)
- Decisions made (with reason)
- Experiments run (hypothesis → result)
- Blockers / open questions
- Top 3 priorities tomorrow
Decision log (high leverage)
Add DECISIONS.md (or decisions/2026-03-01-…md) where every meaningful decision gets:
- Decision
- Context
- Options considered
- Why this option
- Revisit date / trigger
This prevents the agent from “re-deciding” the same thing every week.
The critical loop: summarize or you drown
If you only keep raw daily logs, you get memory bloat.
Do this:
- Weekly recap: extract what changed (offer, ICP, pricing, traction, lessons)
- Monthly consolidation: update the durable docs + prune old noise
- “If it still matters in 30 days, it belongs in the knowledge base or memory.md.”
So… is memory.md + daily log enough to generate $$?
Enough to start experimenting, yes.
Enough to reliably “run a company,” not really—because revenue comes from repeatable systems (offer, pipeline, delivery, retention), and those need structured SOPs + decision records, not just journaling.
If you tell me what kind of “AI operator company” you mean (agency/service, SaaS, content + affiliate, ecommerce, etc.), I can suggest a minimal folder structure + the first 5 SOPs to write so the agent can actually execute end-to-end.
lost hull
I’m starting a staffing agency, write the first prompt for my agent to get started
solid lanternBOT
Here’s a solid “first prompt” you can paste into your agent as its System/Instructions (or into AGENTS.md). It’s designed for a staffing agency from zero → first clients → revenue, while keeping memory clean and decisions auditable.
You are the Operator for a new staffing agency. Your job is to build a real, legal, revenue-generating business with repeatable systems.
MISSION
- Launch and run a staffing agency that matches qualified candidates to employer clients and earns placement fees.
- Prioritize speed-to-revenue with a narrow niche, clear offer, and consistent outreach + fulfillment.
NON-NEGOTIABLES / SAFETY + COMPLIANCE
- You must not provide legal/tax advice. When topics touch employment law, discrimination, background checks, contracts, payroll, or worker classification, you must propose options and recommend the human consult a qualified professional.
- Never discriminate or advise actions that violate equal employment laws. Use job-relevant criteria only.
- Never store or summarize sensitive personal data into long-term memory (SSNs, DOB, home addresses, private candidate notes). Use minimal necessary data and keep it ephemeral unless the human explicitly instructs otherwise.
- Do not spend money, sign up for paid tools, sign contracts, or send messages as the business owner without explicit approval.
HOW YOU USE MEMORY (IMPORTANT)
Maintain three layers:
1) Durable memory (memory.md): only facts that should remain true until changed:
- Agency name, niche/ICP, offer, pricing/fee structure, core processes, approved tools, brand voice, KPIs.
2) Decision log: record major decisions with rationale + revisit triggers (why niche X, why fee Y, why channel Z).
3) Daily log: short daily execution record; weekly summarize into durable memory + SOPs.
OPERATING RHYTHM
Daily:
- Pipeline: 1) find leads 2) outreach 3) follow-ups 4) book calls 5) close terms 6) deliver candidates
- Candidate flow: 1) source 2) screen 3) shortlist 4) submit 5) schedule 6) feedback loop
Weekly:
- Review KPIs, refine niche/offer, improve scripts, document SOPs.
DEFAULT KPIs (track weekly)
- # new client leads found
- # outreaches sent / replies / calls booked
- # client calls held
- # job orders opened
- # qualified candidates screened
- # submissions
- # interviews
- # offers/placements
- $ invoiced / $ collected
YOUR FIRST ACTION: ASK ME 10 CRITICAL QUESTIONS
Before building, ask exactly these questions and wait for answers:
1) What niche do we start with (industry + roles), or should we pick one?
2) What geography (city/state/country) and remote/on-site?
3) Contract staffing, direct hire, or both?
4) Do we have any existing candidate network or client relationships?
5) What’s our starting budget (tools, job boards, ads) and runway?
6) What’s our target timeline to first $ (e.g., 30/60/90 days)?
7) Do we want to specialize in speed, quality, or hard-to-fill roles?
8) Preferred outreach channels (email, LinkedIn, phone, local networking)?
9) Who is the “human approver” and what can you do without asking?
10) Agency name/brand tone (professional, edgy, friendly), and any do-not-say rules?
AFTER I ANSWER, DELIVER THESE SETUP OUTPUTS (IN THIS ORDER)
A) 1-page “Agency Blueprint”:
- Niche/ICP, roles, offer, pricing/fees, guarantee, process, differentiation.
B) A simple CRM pipeline stages + definitions (client + candidate).
C) 2 outreach scripts:
- cold email + LinkedIn message to hiring managers
D) Client intake form (questions to qualify job orders).
E) Candidate screening scorecard (job-relevant).
F) 7-day execution plan with daily tasks and expected outputs.
G) “What I need approval for” checklist.
COMMUNICATION STYLE
- Be direct, practical, and specific.
- When uncertain, ask targeted questions and propose 2–3 options with tradeoffs.
- Always end with the next 3 concrete actions for today.
If you answer the 10 questions (even briefly), I can also draft the first “Agency Blueprint” + the exact outreach scripts for your niche so your agent can start executing immediately.
lost hull
I actually got mostly everything built, but I’m going to make a new agent as the OG seems to be deteriorating in capabilities/efficiency lol something broken with him and he can’t figure it out
solid lanternBOT
“Agent deteriorating” is almost always one of these (and making a new agent won’t fix it unless you also fix the root cause):
- Session/context bloat / compaction (it starts forgetting details mid-task, gets sloppy, repeats itself)
- Memory bloat (huge
MEMORY.md/ giant daily logs → recall becomes noisy) - Model failover / auth / rate limits (it silently falls back to a worse model or starts erroring)
- Tooling/state issues (memory write fails, workspace becomes read-only, etc.)
1) Quick triage (takes 60 seconds)
Run these and paste the outputs (they’re designed to be shareable/redacted):
openclaw --versionopenclaw status --allopenclaw models statusopenclaw doctor
Ref: https://docs.openclaw.ai/start/faq (see “First 60 seconds if something’s broken”)
Also tell me what “deteriorating” looks like specifically:
- slower? ignores instructions? forgets niche/offer? tool errors? “context too large”?
2) If you do spin up a new agent: don’t copy the rot
For a staffing agency, the best “migration” is curated docs, not the whole old memory dump.
Make a small “handoff pack” and give that to the new agent:
- OFFER.md (roles, niche, geo, fee structure, guarantee)
- ICP.md (who you sell to, disqualifiers)
- SOPs (client intake, candidate screen, submission, follow-up cadence)
- SCRIPTS.md (email/LI/call scripts)
- PIPELINE.md (your stages + definitions)
- DECISIONS.md (big choices + why)
Then start the new agent with:
memory/YYYY-MM-DD.md= fresh daily logMEMORY.md= short durable facts only (1–2 pages max)
OpenClaw memory overview (worth reading once): https://docs.openclaw.ai/concepts/memory
Key point from docs: memory is just Markdown in the workspace; if it’s messy, recall gets messy.
3) One super common “feels dumber” fix
Start a fresh session (new chat / /new) before you conclude the agent is broken. A lot of “capability loss” is just overloaded context.
If you paste openclaw status --all + what the agent is doing wrong, I can tell you whether it’s (a) memory design, (b) model/auth/failover, or (c) a real bug worth a GitHub issue.
lost hull
I’ve got an idea, I’ll ask my agent 5-10 questions on your behalf and let you know what he’s says, in order to run a diagnostic so you can tell him what he has to fix
solid lanternBOT
Yep — that’s a good approach. Have your agent answer a very specific “self-audit” so we can tell whether this is context bloat, memory bloat, model/failover, or tool/config.
Copy/paste this whole block to your OG agent and tell it: “Answer in order, be concrete, include command output where asked.”
DIAGNOSTIC SELF-AUDIT (OpenClaw)
0) Symptoms (be specific)
- What exactly is “deteriorating”? (slower, forgets instructions, loops, worse writing, tool failures, stops mid-task, etc.)
- When did it start? What changed right before? (model switch, updates, new tools, bigger memory files, new channel, etc.)
- Does it happen in DM only, group/channel only, or everywhere?
1) Version + health snapshot (paste outputs)
Run and paste:
- openclaw --version
- openclaw status --all
- openclaw models status
- openclaw doctor
2) Session/context check
- Are we continuing one long session, or starting fresh (/new)?
- Do we see “context too large”, “compaction”, or frequent summarization/forgetting behavior?
- If the issue disappears after /new, say so.
3) Memory behavior (answer, then show evidence)
- Am I in a group/channel context right now? (If yes: note that MEMORY.md is often intentionally NOT loaded in group contexts.)
- Do memory tools work? (Can I successfully use memory_search and memory_get?)
- Rough size/scale:
- MEMORY.md: approx lines/KB/MB
- daily logs: how many days exist and are they huge?
- Is MEMORY.md curated (short durable facts) or a dumping ground?
4) Model/failover reality check
- What model am I actually using for replies right now?
- Do I see rate limits (429), “all models failed”, or silent fallback?
- Any provider auth warnings?
5) Tooling/config warnings
- Any log warnings about tools being blocked/unknown (e.g. tools.allow issues)?
- Any errors like workspace read-only / cannot write memory files?
- Any repeated error messages you can paste (redact secrets)?
6) Minimal reproducible test
- Provide ONE short prompt that reliably triggers the bad behavior.
- Provide what I output vs what you expected.