Nobody has built the bot you want.
Verified open-source research on an advisor that nurtures for months and knows when to propose.
Every memory engine stores facts about a person. None stores a sales policy about a relationship.
A search across the whole corpus returns zero open-source projects that model relationship_stage, trust_score, offer_fatigue or upsell_readiness. Every agentic benchmark verified here evaluates a single session. Every piece of sales-agent prior art is a single-session closer.
That gap is the good news. The missing piece is roughly 300 lines of deterministic Python plus a state machine, which makes it your product rather than your integration risk. The research below tells you what to buy for the other eleven layers, and what to refuse.
The 2024 playbook is largely dead code
Letta carries 24,747 stars and no source. Graphiti's last permissively licensed backend was archived. Zep's community edition sits in a folder marked unsupported. Four of the most-cited guardrail tools are archived, one while still pulling 859,000 downloads a month. Star counts and GitHub's pushed_at both lie.
The economics may not clear at your size
Build cost about $24,000, amortised over three years. Breakeven lands near 2,674 customers at a 15 percent lift, 7,910 at 8 percent, and never at 3 percent. Support deflection returns $1,800 to $7,200 a year at 500 to 2,000 customers and arrives in week two. Below roughly 3,000 customers, that is the honest first build.
Four parts
Written for a technical founder and two to four engineers. Every project named carries its real licence, its last commit date, and the reason it might be wrong for you.
Verdict and architecture
The merged architecture, layer by layer, with the runner-up for each and a one-line rule for when to pick it instead. Plus the list of components no open-source project will give you.
The conversational strategy engine
The part that does not exist anywhere in open source. Stages with entry and exit conditions, a readiness score with real feature weights, the conditions that veto an offer outright, and worked transcripts showing internal state beside every turn.
The open-source landscape, by layer
A ranked table per layer: memory, policy engines, prior art, care platforms, channels, grounding, offer timing, signals, guardrails, scheduling, evaluation and runtime. Each carries its real licence and last-commit date.
Roadmap, data model and risk
The day-one schema, four phases with metrics that gate each one, a simulated-customer evaluation plan, the compliance checklist, and what it costs at 5k, 50k and 500k conversations per month.
How this was produced
Thirteen dimensions were researched in parallel, then each was handed to a second agent whose only job was to disbelieve the first: fetch every repository, record the real licence and last commit, and drop anything hallucinated or abandoned. A completeness critic then found five gaps and sent fresh researchers at them, including the cold-start problem, the clienteling alternative, and the fact that the sales science underneath all of this is Anglo-American while the customers here are not.
Three architectures were designed independently and committed to different bets. Three adversarial reviewers scored them through shipability, commercial and trust lenses, and the recommendation is the merge they forced rather than any single proposal.
Research was current as of 15 September 2026. Licences, star counts and maintenance status move; re-verify anything load-bearing before you commit to it.