vortx-ai/emem/emem-referential-drift

emem-referential-drift

Network use. The commands in this skill send JSON to https://emem.dev and receive JSON from it. Where a step says so, they also receive an image or a raster file. https://emem.dev is the service that the MCP server of the plugin connects to. The commands do not execute what they download. This skill ships no scripts. The commands read no environment variable, key or credential. They read only local files that an earlier step saved from https://emem.dev.

Referential drift has two sides. On the words side, "the north field", "plot 14" and a cell64 are one object, or one phrase means two objects. On the values side, agents copy, round and summarise a number until nobody can trace it.

The words side is `emem-shared-identity`. This skill is the values side.

Each call below is a tool of the MCP server of the plugin. Without MCP, send the same JSON body to the REST route that the step names. emem_change_attribution is not in the short tool list that the server shows. The server runs it by name through tools/call. If your host calls only listed tools, use the REST route.

Step 1: get the token for the reading

Each fact that a recall returns has its citation in memory_token. current_by_band names the fact cid of the latest reading for each band. Call emem_recall with this body:

{"cell": "defi.zb493.zezo.zcb35", "bands": ["weather.temperature_2m"]}

Without MCP, send the same body to POST https://emem.dev/v1/recall. Save the result as recall.json and pick the token of the current reading:

jq -r '.current_by_band["weather.temperature_2m"] as $c
       | .facts[] | select(.fact_cid == $c) | .memory_token' recall.json

facts[] is in order from oldest to newest. Thus .facts[0] is the oldest reading, not the current one. On 7 November 2026 at 02:05 UTC this cell held 13 temperature facts, from 23 July to 7 November 2026. The current token changes each time a newer reading lands.

If you have a cell and a fact cid from a different source, build the same token with POST https://emem.dev/v1/memory_token. Send {"cell": ..., "fact_cid": ..., "band": ...} and read memory_token. The result also has the provenance block of the band.

Step 2: resolve the token, quote the signed string

The token is emem:fact:<cell64>:<fact_cid>. Each agent that resolves it gets the same signed object. Call emem_memory_token_resolve, or send the same body to POST https://emem.dev/v1/memory_token/resolve:

{"token": "emem:fact:defi.zb493.zezo.zcb35:nflpddk7zsncywguwjzk5koksseqfyx4jnngkuryrnd4aykqlpfq"}

Read value_verbatim, unit, band and kind in the answer.

On 7 November 2026 this returned {"value_verbatim":"28.0","unit":"degC","band":"weather.temperature_2m","kind":"primary"}. A signed fact does not change, so this token gives the same answer on each day.

Quote value_verbatim. It is the exact decimal string that the responder signed. If you type the JSON number again, you can lose precision.

To resolve many tokens in one call, send {"tokens": [...]} to POST https://emem.dev/v1/memory_token/resolve_many. The limit is 256 tokens (max_tokens). Each entry of items[] has ok and resolution. The value is at items[].resolution.value_verbatim.

Step 3: grade the number before you say it

emem_echo_verify (REST POST https://emem.dev/v1/echo_verify) compares the value that you are about to write with the fact that your token names. Send this body:

{"token": "emem:fact:defi.zb493.zezo.zcb35:nflpddk7zsncywguwjzk5koksseqfyx4jnngkuryrnd4aykqlpfq", "claimed_value": 28.0}

Read matches and drift in the answer.

The check compares the strings first, then the numbers. Results on 7 November 2026:

The 23.9 row uses the token emem:fact:defi.zb493.zezo.zcb35:ofo334twyt6ss3kco6bwljgwy3qiwzqi4grzhog4kwgcpydb6frq. The other rows use the 28.0 token above.

rounded means that you lost precision. wrong means that you have a different number. Use strict when the digits themselves matter, for example in a compliance report. The receipt in the result signs the resolution of the token. It does not sign your claim.

This is the cheapest check in this skill. It finds a failure that you cannot see yourself: a number that you paraphrased some turns before. Run it on each figure that you did not read in this turn from a resolved token.

Step 4: ask why a number moved

Use emem_change_attribution (REST POST https://emem.dev/v1/change_attribution) when the words held and the readout changed between visits. Send this body:

{"cell": "defi.zb493.zezo.zcb35"}

Read observed, split and the keys of terms in the answer.

The result is a ledger of evidence for each term. Each term cites its fact cids:

On 7 November 2026 this call returned all six keys for this cell, with split: null and n_env_evidence_bands: 3.

Two limits apply:

• The result has no numeric split. split is null, and

attribution_note says that the split is roadmap work. Do not report a share such as "60% environmental".

• observed needs a stored embedding. emem.dev no longer computes the

embedding band, so observed is present only where the store already holds a vector. terms.env comes from Sentinel-2 indices and does not need the embedding. Use it first.

On a cold cell the call can return compute_timeout, because it fetches several bands for two visits. This occurred on 28 September 2026. An immediate retry does not help. Recall the index bands of the cell first, wait for the background fetch, then call again. This tool does not accept the MCP task parameter. Only emem_eudr_dds and emem_hunt run as tasks.

Step 5: find where the corpus disagrees

emem_memory_contradictions (REST POST https://emem.dev/v1/memory_contradictions) finds two or more signed values that disagree for the same place, band and time. Send this body:

{"cell": "defi.zb493.zezo.zcb35", "min_severity": 0.2, "limit": 20}

Read the length of contradictions, corpus_scanned and scan_truncated in the answer.

On 7 November 2026 this returned 0 contradictions, with corpus_scanned: 6 and scan_truncated: false.

All fields are optional. cell_prefix limits the scan to a region. band limits it to one measurement. window_unix_s limits it to a period. Without a scope, the scan stops at a cap. On 7 November 2026 a call with only {"limit": 3} scanned 44702 facts and returned scan_truncated: true. Read scan_truncated before you report "none found".

Each entry has a severity from 0 to 1 and a disagreement_scope:

Do not hide a contradiction. Report each signed value and its attester_pubkey_b32. Do not choose a winner, and do not average. No key signed the mean of two values that disagree.

When you must prove the log itself

Sometimes the question is: "Could this responder show a different history to someone else?" For that question, use `emem-transparency-log`.

Common mistakes