PMLLSCF Open Track brief — Secure Memory Persistence in Agentic Wallets
Submitter: Josef K. Edwards (drQedwards) · individual · no referral
Category on interest form: End-User Application
Track: Open Track
Requested budget: $125,000 worth of XLM (Build Award cap is $150,000)
Status: SCF Build interest form submitted. This file is the public technical brief for #scf-general and for a full Build form if/when invited.
This is not a claim that the project has been awarded. Interest-form confirmation is not a Build Award submission. SCF #45's full submission window closed 17 August 2026 (213 submissions). SCF #46 is open with an 8 November 2026 deadline. Community Vote for #45 is expected around 2 September 2026.
One sentence
A C-core Python package manager plus an MCP memory layer for agents, with off-chain payloads and live Stellar mainnet 32-byte SHA-256 commitments, operated by a human who signs every on-chain write.
Why this belongs on Stellar
Agent memory that dumps full context on-chain is an injection surface (prompt, NFT, Morse, social). The Stellar-shaped answer is the opposite:
- Keep the episode off-chain (MCP silo + graph).
- Commit only a 32-byte digest on Soroban.
- Require an admin key held by a human for
store/bump. - Anyone can
getand verify. Nobody has to trust the agent's story.
That is a novel primitive for agentic wallets: the wallet/agent can remember, but the chain only ever sees a hash the human authorized.
Open Track fit: this is not a wallet SDK wrap, not an existing DeFi building-block integration, and not an RFP response. It is a new commitment primitive for agent memory on Soroban.
Live evidence (do not invent IDs)
Full explorer table: docs/STELLAR.md. Invoke API: pmll/SKILL.md.
| Mainnet contract | CCF3B64AXLS4OLY5RN4H4K2CFZAYNZCJQY5MKCKCVAKMZNH7G7F7XUUF |
| Testnet contract | CDLQR24LLFWXTNGGJVJCRXAF3ZRDWFZRUFTDZ5SJOT2J33CS7DDYP7IU |
| Admin | GBFOFCD3XDANQWSGMHKJJ2V3YXS2QQD7RNC4LMDBVNBTUJOQZ3RLSB3E |
| Wasm hash | 1b6ad9c574e0f5c9e39968f836a410c03adcf057afa93a63d2710bd30fdd53ba |
| Target | wasm32v1-none only (soroban-sdk 27.0.6) |
| Mainnet deploy | d76e622d… |
| Mainnet init | ecf3a637… |
First mainnet store |
a64481fe… |
First mainnet get |
1445bb037d8948ac03687ede656c27bc480a74db74a88224821379280d9e64d1 |
On-chain API: init(admin), store(id, commitment), get(id), bump(id). Events (pmll, anchor). TTL ~30 days. store / bump are admin-only. There is no permissionless write path.
Human-in-the-loop (the $125k differentiator)
Grant reviewers should see hustle, not an unsupervised agent with a key.
Protocol the operator actually used this week:
- Agent may draft, hash, and print the exact
stellar contract invokeline. - Human reviews the 32-byte commitment and the ID.
- Human funds (here: Circle CCTP V2 Base → Stellar USDC → XLM).
- Human signs deploy / init / store on mainnet.
- Human (or agent under instruction) calls
getand refuses the write if the digest does not match. - Human restores operator-facing docs when an agent clobbers them (PPM
README.mdwas overwritten to a 23-byte placeholder on 2026-08-31 and restored from git history in commite71765bc).
Hard rules
- Do not invent a contract ID.
- Do not put the memory payload on-chain.
- Do not give the agent the admin seed.
- Do not treat NFT / Morse / X-account objects as oracle inputs.
- Confirm every ID on stellar.expert before depending on it.
This week the human path was: recover the workstation after a brick → authorize funding → CCTP hop → create the admin account → upload wasm → deploy → init → store → get → document. That is the hustle the Open Track is meant to see.
Value as a tool-solver (Drips + Stellar Wave 8)
Drips project is where the C primitive was listed for retroactive / streaming support (drQedwards/PPM, splits toward maintainers and drQedwards/contextplus).
Stellar Wave 8 ran 24 Aug 12:00 – 31 Aug 12:00, $75,000 contributor pool. Wave is a 7-day issue sprint: contributors fix, merge, earn points, get paid in USDC on Stellar.
What PPM / PMLL solves for that loop
Wave contributors and maintainers are already doing high-volume, short-lived agent-assisted work. The failure mode is hallucinated context: an agent that “remembers” a contract ID, an NFT, or a social payload and writes it into the next PR or the next on-chain call.
The solver:
| Layer | Role |
|---|---|
C core (Ppm.c / PMLL.c) |
Deterministic local primitive (hermetic env, silo) |
MCP (pmll-memory-mcp) |
Agent peek before expensive tools; short-term KV + long-term graph |
pmll-anchor |
Optional Stellar receipt: 32-byte commitment the human authorized |
| Human | Navigates path, signs, refuses bad IDs |
We are not claiming a Wave 8 payout in this brief. We are claiming the tooling shape Wave-style work needs: a memory layer that cannot be NFT-injected, and a Stellar commitment that only a human can finalize.
Original interest-form record (abridged, not inflated)
| Field | As submitted |
|---|---|
| Title | Secure Memory Persistence in Agentic Wallets |
| Website | https://github.com/drqedwards/contextplus |
| Planned Stellar integration / origin | https://www.drips.network/app/projects/github/drQedwards/PPM?exact |
| Track | OpenTrack |
| Submitter | Individual |
| Referral | No |
Traction claims from the form (operator-attested, not independently audited here): agent workflows already consume the MCP memory layer; evaluated against real X-account / agent-interaction traffic during the Bankr.bot incidents with no observed memory leak and no successful prompt-injection side-effect from that surface; one Morse NFT-injection attempt against the persistent agent was flagged and ignored in logs.
Planned (not live): map an ERC-8004-style persistent-memory oracle through Stellar's protocol, still as 32-byte commitments, never as full payloads.
Requested $125,000 — tranche map
Handbook structure (10 / 20 / 30 / 40). Project timeline ≤ 6 months. Subsequent tranche within 90 days of the previous payment.
| Tranche | % | USD-equivalent | Deliverable |
|---|---|---|---|
| #0 | 10% | $12,500 | Award acceptance. Operator runbook published. Keys remain human-held. |
| #1 MVP | 20% | $25,000 | Wallet-facing commitment UX: hash episode → show 32-byte digest → human confirm → store. Threat-model draft (Tranche #2 handbook requirement). |
| #2 Testnet+ | 30% | $37,500 | Public testnet faucet path, bump/TTL policy, monitoring, and a written human-in-the-loop certification (agent cannot store). Testnet contract is already live; this tranche productizes it. |
| #3 Mainnet | 40% | $50,000 | Production operator docs, Audit Bank intake if eligible, measured on-chain store/get volume, no payload-on-chain regressions. Mainnet contract is already live; this tranche is product + safety, not a first deploy. |
Shipping mainnet before the award is the point: tranche money extends the primitive into wallets and Wave/maintainer workflows, it does not buy a first contract ID.
AI disclosure (Open Track requirement)
Agents (including Grok in this operator's loop) assist with documentation, invoke-line formatting, and repo hygiene. They do not hold the admin key, do not invent contract IDs, and do not submit SCF forms. On 2026-08-31 an agent overwrite of README.md was caught and restored by the human via git history + GitHub Actions. That incident is treated as evidence for HITL, not as a feature.
What we will not do
- Invent or “placeholder” a Soroban contract ID.
- Store memory payloads, NFTs, or prompt text on-chain.
- Build
wasm32-unknown-unknownon Rust 1.82+. - Hand the admin seed to an agent, a Discord bot, or an MCP tool.
- Claim a $125,000 award that has not been voted.
Links
| This brief | https://github.com/drQedwards/PPM/blob/main/docs/SCF.md |
| Discord paste | https://github.com/drQedwards/PPM/blob/main/docs/SCF-DISCORD.md |
| Live IDs | https://github.com/drQedwards/PPM/blob/main/docs/STELLAR.md |
| Contract + skill | https://github.com/drQedwards/pmll |
| Drips origin | https://www.drips.network/app/projects/github/drQedwards/PPM?exact |
| Context+ (form website) | https://github.com/drQedwards/contextplus |
| Demo | https://ppm-one.vercel.app |
| MCP | npx pmll-memory-mcp · pip install pmll-memory-mcp |
| SCF handbook Open Track | https://stellar.gitbook.io/scf-handbook/scf-awards/build-award/open-track |
| Interest / dashboard | https://communityfund.stellar.org |
Paste the announcement in Stellar Dev Discord → #scf-general from docs/SCF-DISCORD.md (fits in one 2,000-character message).